[client] Merge main into peer event bus refactor

client/internal/acl/dispatch_test.go

@@ -0,0 +1,190 @@
+package acl
+
+import (
+	"net/netip"
+	"sync"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"go.uber.org/mock/gomock"
+
+	"github.com/netbirdio/netbird/client/firewall"
+	fwmgr "github.com/netbirdio/netbird/client/firewall/manager"
+	"github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+	"github.com/netbirdio/netbird/client/internal/acl/mocks"
+	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
+)
+
+// TestNetworkZeroPrefixIsRoute guards the route-vs-peer dispatch
+// invariant: the backends classify a rule as a peer rule purely by the
+// absence of a destination (neither prefix nor set). A default route
+// (0.0.0.0/0 or ::/0) is a valid prefix and must therefore classify as
+// a route, not collapse into the peer path.
+func TestNetworkZeroPrefixIsRoute(t *testing.T) {
+	for _, p := range []string{"0.0.0.0/0", "::/0", "10.0.0.0/8"} {
+		n := fwmgr.Network{Prefix: netip.MustParsePrefix(p)}
+		assert.True(t, n.IsPrefix(), "%s must report IsPrefix", p)
+		assert.True(t, n.IsPrefix() || n.IsSet(), "%s must classify as a route", p)
+	}
+
+	// A zero-value Network is the only peer-rule shape.
+	var empty fwmgr.Network
+	assert.False(t, empty.IsPrefix(), "zero Network must not be a prefix")
+	assert.False(t, empty.IsSet(), "zero Network must not be a set")
+}
+
+// TestDetermineDestinationAlwaysRoute verifies determineDestination
+// never yields an empty Network for a valid route rule: every branch
+// (static prefix, default route, dynamic with/without domains, with and
+// without a local resolver) produces a destination that classifies as a
+// route. If this regresses, a route rule would be dispatched down the
+// peer path, which matches on source only.
+func TestDetermineDestinationAlwaysRoute(t *testing.T) {
+	v4 := []netip.Prefix{netip.MustParsePrefix("10.0.0.0/24")}
+	v6 := []netip.Prefix{netip.MustParsePrefix("2001:db8::/48")}
+
+	cases := []struct {
+		name     string
+		rule     *mgmProto.RouteFirewallRule
+		resolver bool
+		sources  []netip.Prefix
+	}{
+		{"static prefix", &mgmProto.RouteFirewallRule{Destination: "192.168.0.0/16"}, false, v4},
+		{"static default route", &mgmProto.RouteFirewallRule{Destination: "0.0.0.0/0"}, false, v4},
+		{"dynamic with domains + resolver", &mgmProto.RouteFirewallRule{IsDynamic: true, Domains: []string{"example.com"}}, true, v4},
+		{"dynamic no domains + resolver (v4)", &mgmProto.RouteFirewallRule{IsDynamic: true}, true, v4},
+		{"dynamic no domains + resolver (v6)", &mgmProto.RouteFirewallRule{IsDynamic: true}, true, v6},
+		{"dynamic + no local resolver (v4)", &mgmProto.RouteFirewallRule{IsDynamic: true}, false, v4},
+		{"dynamic + no local resolver (v6)", &mgmProto.RouteFirewallRule{IsDynamic: true}, false, v6},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			dest, err := determineDestination(tc.rule, tc.resolver, tc.sources)
+			require.NoError(t, err)
+			assert.True(t, dest.IsPrefix() || dest.IsSet(),
+				"destination must classify as a route, got empty Network")
+		})
+	}
+}
+
+// countingFirewall wraps a real firewall.Manager and counts filter-rule
+// add/delete calls so a test can assert how many backing rules the acl
+// manager actually creates and tears down.
+type countingFirewall struct {
+	fwmgr.Manager
+	mu       sync.Mutex
+	addCalls int
+	dels     int
+	ruleIDs  map[fwmgr.RuleID]struct{}
+}
+
+// distinctRules returns the number of distinct backing rules the
+// backend produced. Because the backend dedups identical content,
+// repeated AddFilterRule calls for the same rule resolve to one id.
+func (f *countingFirewall) distinctRules() int {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return len(f.ruleIDs)
+}
+
+func (f *countingFirewall) AddFilterRule(id []byte, sources []netip.Prefix, destination fwmgr.Network, proto fwmgr.Protocol, sPort, dPort *fwmgr.Port, action fwmgr.Action) (fwmgr.Rule, error) {
+	rule, err := f.Manager.AddFilterRule(id, sources, destination, proto, sPort, dPort, action)
+	if err == nil {
+		f.mu.Lock()
+		f.addCalls++
+		if f.ruleIDs == nil {
+			f.ruleIDs = make(map[fwmgr.RuleID]struct{})
+		}
+		if rule != nil {
+			f.ruleIDs[rule.ID()] = struct{}{}
+		}
+		f.mu.Unlock()
+	}
+	return rule, err
+}
+
+func (f *countingFirewall) DeleteFilterRule(r fwmgr.Rule) error {
+	err := f.Manager.DeleteFilterRule(r)
+	if err == nil {
+		f.mu.Lock()
+		f.dels++
+		delete(f.ruleIDs, r.ID())
+		f.mu.Unlock()
+	}
+	return err
+}
+
+func newCountingACL(t *testing.T) (*DefaultManager, *countingFirewall, func()) {
+	t.Helper()
+	t.Setenv("NB_WG_KERNEL_DISABLED", "true")
+	t.Setenv(firewall.EnvForceUserspaceFirewall, "true")
+
+	ctrl := gomock.NewController(t)
+	ifaceMock := mocks.NewMockIFaceMapper(ctrl)
+	ifaceMock.EXPECT().IsUserspaceBind().Return(true).AnyTimes()
+	ifaceMock.EXPECT().SetFilter(gomock.Any())
+	network := netip.MustParsePrefix("172.0.0.1/32")
+	ifaceMock.EXPECT().Name().Return("lo").AnyTimes()
+	ifaceMock.EXPECT().Address().Return(wgaddr.Address{IP: network.Addr(), Network: network}).AnyTimes()
+	ifaceMock.EXPECT().GetWGDevice().Return(nil).AnyTimes()
+
+	realFW, err := firewall.NewFirewall(ifaceMock, nil, flowLogger, false, iface.DefaultMTU)
+	require.NoError(t, err)
+
+	fw := &countingFirewall{Manager: realFW}
+	cleanup := func() {
+		require.NoError(t, realFW.Close(nil))
+		ctrl.Finish()
+	}
+	return NewDefaultManager(fw), fw, cleanup
+}
+
+// TestDuplicateContentPoliciesShareOneRule verifies the dedup contract
+// the backends rely on: two policies that authorize an identical flow
+// (same selector and sources) collapse to a single backing firewall
+// rule, and that rule survives until BOTH policies are gone. This is
+// why the backend can dedup on add without refcounting on delete: the
+// acl manager's pair key matches the backend's content key, so add and
+// delete stay balanced per content key across full-state reapplies.
+func TestDuplicateContentPoliciesShareOneRule(t *testing.T) {
+	acl, fw, cleanup := newCountingACL(t)
+	defer cleanup()
+
+	ruleA := &mgmProto.FirewallRule{
+		PolicyID:  []byte("policy-A"),
+		PeerIP:    "10.0.0.1", //nolint:staticcheck
+		Direction: mgmProto.RuleDirection_IN,
+		Action:    mgmProto.RuleAction_ACCEPT,
+		Protocol:  mgmProto.RuleProtocol_TCP,
+		Port:      "443",
+	}
+	ruleB := &mgmProto.FirewallRule{
+		PolicyID:  []byte("policy-B"),
+		PeerIP:    "10.0.0.1", //nolint:staticcheck
+		Direction: mgmProto.RuleDirection_IN,
+		Action:    mgmProto.RuleAction_ACCEPT,
+		Protocol:  mgmProto.RuleProtocol_TCP,
+		Port:      "443",
+	}
+
+	// Both policies present: identical content collapses to one rule.
+	acl.ApplyFiltering(&mgmProto.NetworkMap{FirewallRules: []*mgmProto.FirewallRule{ruleA, ruleB}, FirewallRulesIsEmpty: false}, false)
+	assert.Equal(t, 1, fw.distinctRules(), "identical-content policies must produce one backing rule")
+	assert.Equal(t, 1, len(acl.peerRulesPairs), "one content key, one pair")
+
+	// Drop policy A only: the shared rule is still authorized by B, so
+	// nothing is deleted.
+	acl.ApplyFiltering(&mgmProto.NetworkMap{FirewallRules: []*mgmProto.FirewallRule{ruleB}, FirewallRulesIsEmpty: false}, false)
+	assert.Equal(t, 1, fw.distinctRules(), "no new backing rule on reapply")
+	assert.Equal(t, 0, fw.dels, "rule must survive while any policy still authorizes it")
+	assert.Equal(t, 1, len(acl.peerRulesPairs))
+
+	// Drop policy B too: now the content key has no authorizer and the
+	// single backing rule is removed exactly once.
+	acl.ApplyFiltering(&mgmProto.NetworkMap{FirewallRules: nil, FirewallRulesIsEmpty: true}, false)
+	assert.Equal(t, 1, fw.dels, "rule removed once when last policy is gone")
+	assert.Equal(t, 0, len(acl.peerRulesPairs))
+}

client/internal/acl/grouping_test.go

@@ -0,0 +1,318 @@
+package acl
+
+import (
+	"errors"
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"go.uber.org/mock/gomock"
+
+	"github.com/netbirdio/netbird/client/firewall"
+	fwmgr "github.com/netbirdio/netbird/client/firewall/manager"
+	"github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+	"github.com/netbirdio/netbird/client/internal/acl/mocks"
+	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
+	"github.com/netbirdio/netbird/shared/netiputil"
+)
+
+// TestGroupPeerRulesPolicyIDSeparates verifies that two FirewallRules
+// with identical selectors but different PolicyIDs do NOT get merged
+// into one group, so each policy's sources merge under its own
+// attribution id. (Identical-content groups may still dedup to one
+// backing rule at the backend; see TestDuplicateContentPoliciesShareOneRule.)
+func TestGroupPeerRulesPolicyIDSeparates(t *testing.T) {
+	rules := []*mgmProto.FirewallRule{
+		{
+			PolicyID:  []byte("policy-A"),
+			PeerIP:    "10.0.0.1",
+			Direction: mgmProto.RuleDirection_IN,
+			Action:    mgmProto.RuleAction_ACCEPT,
+			Protocol:  mgmProto.RuleProtocol_TCP,
+			Port:      "443",
+		},
+		{
+			PolicyID:  []byte("policy-B"),
+			PeerIP:    "10.0.0.1",
+			Direction: mgmProto.RuleDirection_IN,
+			Action:    mgmProto.RuleAction_ACCEPT,
+			Protocol:  mgmProto.RuleProtocol_TCP,
+			Port:      "443",
+		},
+	}
+
+	groups, denyErr, err := groupPeerRules(rules)
+	require.NoError(t, denyErr)
+	require.NoError(t, err)
+	require.Len(t, groups, 2, "rules with different PolicyIDs must produce separate groups")
+}
+
+// TestGroupPeerRulesFamilySeparates verifies that v4 and v6 rules
+// belonging to the same policy don't merge.
+func TestGroupPeerRulesFamilySeparates(t *testing.T) {
+	rules := []*mgmProto.FirewallRule{
+		{
+			PolicyID:  []byte("policy-A"),
+			PeerIP:    "10.0.0.1",
+			Direction: mgmProto.RuleDirection_IN,
+			Action:    mgmProto.RuleAction_ACCEPT,
+			Protocol:  mgmProto.RuleProtocol_TCP,
+			Port:      "443",
+		},
+		{
+			PolicyID:  []byte("policy-A"),
+			PeerIP:    "2001:db8::1",
+			Direction: mgmProto.RuleDirection_IN,
+			Action:    mgmProto.RuleAction_ACCEPT,
+			Protocol:  mgmProto.RuleProtocol_TCP,
+			Port:      "443",
+		},
+	}
+
+	groups, denyErr, err := groupPeerRules(rules)
+	require.NoError(t, denyErr)
+	require.NoError(t, err)
+	require.Len(t, groups, 2, "rules of different families must produce separate groups")
+
+	var sawV4, sawV6 bool
+	for _, g := range groups {
+		require.Len(t, g.sources, 1)
+		if g.sources[0].Addr().Is4() {
+			sawV4 = true
+		}
+		if g.sources[0].Addr().Is6() {
+			sawV6 = true
+		}
+	}
+	assert.True(t, sawV4 && sawV6)
+}
+
+// TestGroupPeerRulesSplitsMixedFamilySingleRule verifies that a single
+// FirewallRule carrying both v4 and v6 source prefixes is split into one
+// group per family. Each backend keys a rule to a single family, so a
+// group whose sources span families would mismatch the other family's
+// sources. mgmt normally emits one rule per family; this guards against
+// a mixed-family rule slipping through.
+func TestGroupPeerRulesSplitsMixedFamilySingleRule(t *testing.T) {
+	srcs := [][]byte{
+		netiputil.EncodeAddr(netip.MustParseAddr("10.0.0.1")),
+		netiputil.EncodeAddr(netip.MustParseAddr("2001:db8::1")),
+		netiputil.EncodeAddr(netip.MustParseAddr("10.0.0.2")),
+		netiputil.EncodeAddr(netip.MustParseAddr("2001:db8::2")),
+	}
+	rules := []*mgmProto.FirewallRule{
+		{
+			PolicyID:       []byte("policy-A"),
+			SourcePrefixes: srcs,
+			Direction:      mgmProto.RuleDirection_IN,
+			Action:         mgmProto.RuleAction_ACCEPT,
+			Protocol:       mgmProto.RuleProtocol_TCP,
+			Port:           "443",
+		},
+	}
+
+	groups, denyErr, err := groupPeerRules(rules)
+	require.NoError(t, denyErr)
+	require.NoError(t, err)
+	require.Len(t, groups, 2, "mixed-family sources in one rule must split into two groups")
+
+	for _, g := range groups {
+		require.Len(t, g.sources, 2)
+		v6 := prefixIsV6(g.sources[0])
+		for _, s := range g.sources {
+			assert.Equal(t, v6, prefixIsV6(s), "every source in a group must share one family")
+		}
+	}
+}
+
+// TestGroupPeerRulesMergesSameSelector verifies that rules sharing
+// every distinguishing field (policy, family, direction, action,
+// proto, port) collapse into a single multi-source group.
+func TestGroupPeerRulesMergesSameSelector(t *testing.T) {
+	mk := func(peerIP string) *mgmProto.FirewallRule {
+		return &mgmProto.FirewallRule{
+			PolicyID:  []byte("policy-A"),
+			PeerIP:    peerIP, //nolint:staticcheck
+			Direction: mgmProto.RuleDirection_IN,
+			Action:    mgmProto.RuleAction_ACCEPT,
+			Protocol:  mgmProto.RuleProtocol_TCP,
+			Port:      "443",
+		}
+	}
+	rules := []*mgmProto.FirewallRule{mk("10.0.0.1"), mk("10.0.0.2"), mk("10.0.0.3")}
+
+	groups, denyErr, err := groupPeerRules(rules)
+	require.NoError(t, denyErr)
+	require.NoError(t, err)
+	require.Len(t, groups, 1)
+	require.Len(t, groups[0].sources, 3)
+}
+
+// TestGroupPeerRulesPortSeparates verifies that PortInfo is part of the
+// selector key: rules differing only in port must not merge, and a
+// single port must not merge with a range. A regression dropping the
+// port from the key would collapse rules for different ports into one.
+func TestGroupPeerRulesPortSeparates(t *testing.T) {
+	mkPort := func(peerIP string, port uint32) *mgmProto.FirewallRule {
+		return &mgmProto.FirewallRule{
+			PolicyID:  []byte("policy-A"),
+			PeerIP:    peerIP, //nolint:staticcheck
+			Direction: mgmProto.RuleDirection_IN,
+			Action:    mgmProto.RuleAction_ACCEPT,
+			Protocol:  mgmProto.RuleProtocol_TCP,
+			PortInfo:  &mgmProto.PortInfo{PortSelection: &mgmProto.PortInfo_Port{Port: port}},
+		}
+	}
+
+	groups, denyErr, err := groupPeerRules([]*mgmProto.FirewallRule{
+		mkPort("10.0.0.1", 80), mkPort("10.0.0.2", 80), mkPort("10.0.0.3", 443),
+	})
+	require.NoError(t, denyErr)
+	require.NoError(t, err)
+	require.Len(t, groups, 2, "rules on different ports must not merge")
+
+	rangeRule := &mgmProto.FirewallRule{
+		PolicyID:  []byte("policy-A"),
+		PeerIP:    "10.0.0.4", //nolint:staticcheck
+		Direction: mgmProto.RuleDirection_IN,
+		Action:    mgmProto.RuleAction_ACCEPT,
+		Protocol:  mgmProto.RuleProtocol_TCP,
+		PortInfo:  &mgmProto.PortInfo{PortSelection: &mgmProto.PortInfo_Range_{Range: &mgmProto.PortInfo_Range{Start: 80, End: 90}}},
+	}
+	groups, denyErr, err = groupPeerRules([]*mgmProto.FirewallRule{mkPort("10.0.0.1", 80), rangeRule})
+	require.NoError(t, denyErr)
+	require.NoError(t, err)
+	require.Len(t, groups, 2, "a single port and a range must not merge")
+}
+
+// TestGroupPeerRulesUsesSourcePrefixesWhenPresent verifies that the
+// new sourcePrefixes wire field is consumed and produces a
+// multi-source group in one shot (no client-side merging needed).
+func TestGroupPeerRulesUsesSourcePrefixesWhenPresent(t *testing.T) {
+	srcs := [][]byte{
+		netiputil.EncodeAddr(netip.MustParseAddr("10.0.0.1")),
+		netiputil.EncodeAddr(netip.MustParseAddr("10.0.0.2")),
+		netiputil.EncodeAddr(netip.MustParseAddr("10.0.0.3")),
+	}
+	rules := []*mgmProto.FirewallRule{
+		{
+			PolicyID:       []byte("policy-A"),
+			SourcePrefixes: srcs,
+			Direction:      mgmProto.RuleDirection_IN,
+			Action:         mgmProto.RuleAction_ACCEPT,
+			Protocol:       mgmProto.RuleProtocol_TCP,
+			Port:           "443",
+		},
+	}
+
+	groups, denyErr, err := groupPeerRules(rules)
+	require.NoError(t, denyErr)
+	require.NoError(t, err)
+	require.Len(t, groups, 1)
+	require.Len(t, groups[0].sources, 3)
+}
+
+// TestGroupPeerRulesActionSeparates verifies the obvious: accept
+// and drop rules with the same selector don't merge.
+func TestGroupPeerRulesActionSeparates(t *testing.T) {
+	rules := []*mgmProto.FirewallRule{
+		{
+			PolicyID:  []byte("policy-A"),
+			PeerIP:    "10.0.0.1",
+			Direction: mgmProto.RuleDirection_IN,
+			Action:    mgmProto.RuleAction_ACCEPT,
+			Protocol:  mgmProto.RuleProtocol_TCP,
+			Port:      "443",
+		},
+		{
+			PolicyID:  []byte("policy-A"),
+			PeerIP:    "10.0.0.1",
+			Direction: mgmProto.RuleDirection_IN,
+			Action:    mgmProto.RuleAction_DROP,
+			Protocol:  mgmProto.RuleProtocol_TCP,
+			Port:      "443",
+		},
+	}
+
+	groups, denyErr, err := groupPeerRules(rules)
+	require.NoError(t, denyErr)
+	require.NoError(t, err)
+	require.Len(t, groups, 2)
+}
+
+// failingDeleteFirewall wraps a real firewall.Manager and forces the
+// next N DeleteFilterRule calls to fail. Used to verify that the acl
+// manager retains rules whose deletion was rejected by the backend,
+// so they get retried on the next ApplyFiltering pass instead of
+// becoming orphans.
+type failingDeleteFirewall struct {
+	fwmgr.Manager
+	failCount int
+}
+
+func (f *failingDeleteFirewall) DeleteFilterRule(r fwmgr.Rule) error {
+	if f.failCount > 0 {
+		f.failCount--
+		return errors.New("simulated delete failure")
+	}
+	return f.Manager.DeleteFilterRule(r)
+}
+
+// TestApplyFilteringRetainsRulesOnDeleteFailure verifies that a
+// transient DeleteFilterRule error doesn't make the acl manager
+// forget about a rule. The rule must remain in peerRulesPairs so the
+// next ApplyFiltering pass attempts the delete again.
+func TestApplyFilteringRetainsRulesOnDeleteFailure(t *testing.T) {
+	t.Setenv("NB_WG_KERNEL_DISABLED", "true")
+	t.Setenv(firewall.EnvForceUserspaceFirewall, "true")
+
+	ctrl := gomock.NewController(t)
+	defer ctrl.Finish()
+
+	ifaceMock := mocks.NewMockIFaceMapper(ctrl)
+	ifaceMock.EXPECT().IsUserspaceBind().Return(true).AnyTimes()
+	ifaceMock.EXPECT().SetFilter(gomock.Any())
+	network := netip.MustParsePrefix("172.0.0.1/32")
+	ifaceMock.EXPECT().Name().Return("lo").AnyTimes()
+	ifaceMock.EXPECT().Address().Return(wgaddr.Address{IP: network.Addr(), Network: network}).AnyTimes()
+	ifaceMock.EXPECT().GetWGDevice().Return(nil).AnyTimes()
+
+	realFW, err := firewall.NewFirewall(ifaceMock, nil, flowLogger, false, iface.DefaultMTU)
+	require.NoError(t, err)
+	defer func() { require.NoError(t, realFW.Close(nil)) }()
+
+	fw := &failingDeleteFirewall{Manager: realFW}
+	acl := NewDefaultManager(fw)
+
+	// First pass: install a rule.
+	netmap1 := &mgmProto.NetworkMap{
+		FirewallRules: []*mgmProto.FirewallRule{
+			{
+				PolicyID:  []byte("policy-A"),
+				PeerIP:    "10.0.0.1",
+				Direction: mgmProto.RuleDirection_IN,
+				Action:    mgmProto.RuleAction_DROP,
+				Protocol:  mgmProto.RuleProtocol_TCP,
+				Port:      "22",
+			},
+		},
+		FirewallRulesIsEmpty: false,
+	}
+	acl.ApplyFiltering(netmap1, false)
+	require.Equal(t, 1, len(acl.peerRulesPairs), "rule should be installed")
+
+	// Second pass: remove the rule from the map. The backend will
+	// fail the delete; the acl manager must retain the rule.
+	fw.failCount = 1
+	netmap2 := &mgmProto.NetworkMap{FirewallRules: nil, FirewallRulesIsEmpty: true}
+	acl.ApplyFiltering(netmap2, false)
+	require.Equal(t, 1, len(acl.peerRulesPairs),
+		"rule must be retained when DeleteFilterRule fails so it gets retried")
+
+	// Third pass: same map, backend no longer fails. The rule
+	// should now succeed in being removed.
+	acl.ApplyFiltering(netmap2, false)
+	require.Equal(t, 0, len(acl.peerRulesPairs), "retry should succeed")
+}

client/internal/acl/id/id.go

@@ -5,18 +5,18 @@ import (
 	"encoding/hex"
 	"fmt"
 	"net/netip"
+	"slices"
 	"strconv"
 
 	"github.com/netbirdio/netbird/client/firewall/manager"
 )
 
-type RuleID string
+// RuleID aliases manager.RuleID so existing nbid.RuleID references
+// keep working while the canonical type lives in the firewall package.
+type RuleID = manager.RuleID
 
-func (r RuleID) ID() string {
-	return string(r)
-}
-
-func GenerateRouteRuleKey(
+// GenerateRuleID returns a deterministic content hash identifying a filter rule.
+func GenerateRuleID(
 	sources []netip.Prefix,
 	destination manager.Network,
 	proto manager.Protocol,
@@ -24,6 +24,7 @@ func GenerateRouteRuleKey(
 	dPort *manager.Port,
 	action manager.Action,
 ) RuleID {
+	sources = slices.Clone(sources)
 	manager.SortPrefixes(sources)
 
 	h := sha256.New()

client/internal/acl/legacy_fallback_test.go

@@ -0,0 +1,75 @@
+package acl
+
+import (
+	"net/netip"
+	"sync"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"go.uber.org/mock/gomock"
+
+	"github.com/netbirdio/netbird/client/firewall"
+	fwmgr "github.com/netbirdio/netbird/client/firewall/manager"
+	"github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+	"github.com/netbirdio/netbird/client/internal/acl/mocks"
+	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
+)
+
+// sourcesRecordingFirewall wraps a real firewall.Manager and records
+// the source prefixes of every AddFilterRule call.
+type sourcesRecordingFirewall struct {
+	fwmgr.Manager
+	mu      sync.Mutex
+	sources [][]netip.Prefix
+}
+
+func (f *sourcesRecordingFirewall) AddFilterRule(id []byte, sources []netip.Prefix, destination fwmgr.Network, proto fwmgr.Protocol, sPort, dPort *fwmgr.Port, action fwmgr.Action) (fwmgr.Rule, error) {
+	f.mu.Lock()
+	f.sources = append(f.sources, sources)
+	f.mu.Unlock()
+	return f.Manager.AddFilterRule(id, sources, destination, proto, sPort, dPort, action)
+}
+
+// TestLegacyManagementFallbackUsesMatchAnySources verifies the
+// allow-all fallback for old management servers (empty FirewallRules
+// without the FirewallRulesIsEmpty flag) reaches the firewall as /0
+// match-any sources. The fallback rule carries PeerIP 0.0.0.0; if that
+// were converted to a host prefix (0.0.0.0/32) it would match nothing
+// and all peer traffic would be dropped.
+func TestLegacyManagementFallbackUsesMatchAnySources(t *testing.T) {
+	t.Setenv("NB_WG_KERNEL_DISABLED", "true")
+	t.Setenv(firewall.EnvForceUserspaceFirewall, "true")
+
+	ctrl := gomock.NewController(t)
+	defer ctrl.Finish()
+
+	ifaceMock := mocks.NewMockIFaceMapper(ctrl)
+	ifaceMock.EXPECT().IsUserspaceBind().Return(true).AnyTimes()
+	ifaceMock.EXPECT().SetFilter(gomock.Any())
+	network := netip.MustParsePrefix("172.0.0.1/32")
+	ifaceMock.EXPECT().Name().Return("lo").AnyTimes()
+	ifaceMock.EXPECT().Address().Return(wgaddr.Address{IP: network.Addr(), Network: network}).AnyTimes()
+	ifaceMock.EXPECT().GetWGDevice().Return(nil).AnyTimes()
+
+	realFW, err := firewall.NewFirewall(ifaceMock, nil, flowLogger, false, iface.DefaultMTU)
+	require.NoError(t, err)
+	defer func() { require.NoError(t, realFW.Close(nil)) }()
+
+	fw := &sourcesRecordingFirewall{Manager: realFW}
+	acl := NewDefaultManager(fw)
+
+	// Old management: no rules and no FirewallRulesIsEmpty flag.
+	acl.ApplyFiltering(&mgmProto.NetworkMap{FirewallRules: nil, FirewallRulesIsEmpty: false}, false)
+
+	fw.mu.Lock()
+	defer fw.mu.Unlock()
+	require.NotEmpty(t, fw.sources, "legacy fallback must install at least one allow-all rule")
+	for _, sources := range fw.sources {
+		require.NotEmpty(t, sources)
+		for _, p := range sources {
+			assert.Equal(t, 0, p.Bits(), "legacy fallback source %s must be a /0 match-any prefix", p)
+		}
+	}
+}

client/internal/acl/manager.go

@@ -1,8 +1,6 @@
 package acl
 
 import (
-	"crypto/md5"
-	"encoding/hex"
 	"errors"
 	"fmt"
 	"net/netip"
@@ -24,6 +22,10 @@ import (
 
 var ErrSourceRangesEmpty = errors.New("sources range is empty")
 
+// ErrNoRuleReturned is returned when the firewall backend reports success
+// from AddFilterRule but yields no rule to track.
+var ErrNoRuleReturned = errors.New("backend returned no rule")
+
 // Manager is a ACL rules manager
 type Manager interface {
 	ApplyFiltering(networkMap *mgmProto.NetworkMap, dnsRouteFeatureFlag bool)
@@ -32,19 +34,48 @@ type Manager interface {
 // DefaultManager uses firewall manager to handle
 type DefaultManager struct {
 	firewall           firewall.Manager
-	ipsetCounter       int
 	peerRulesPairs     map[id.RuleID][]firewall.Rule
-	routeRules         map[id.RuleID]struct{}
+	routeRules         map[id.RuleID]firewall.Rule
 	previousConfigHash uint64
 	hasAppliedConfig   bool
 	mutex              sync.Mutex
 }
 
+// peerRuleGroup collapses a set of single-source FirewallRules sharing
+// the same selector into one multi-source rule to push to the backend.
+type peerRuleGroup struct {
+	direction mgmProto.RuleDirection
+	action    mgmProto.RuleAction
+	protocol  mgmProto.RuleProtocol
+	port      *mgmProto.PortInfo
+	// legacyPort is used only when PortInfo is empty (old management).
+	legacyPort string
+	policyID   []byte
+	sources    []netip.Prefix
+}
+
+// peerRuleKey is the comparable selector that decides which single-source
+// rules merge into one group. Rules with an equal key collapse into one
+// multi-source backend rule. PortInfo is flattened into its scalar fields
+// so the key compares by value; policyID keeps policies separate so two
+// policies authorizing different peers don't merge under one attribution.
+type peerRuleKey struct {
+	v6         bool
+	policyID   string
+	direction  mgmProto.RuleDirection
+	action     mgmProto.RuleAction
+	protocol   mgmProto.RuleProtocol
+	legacyPort string
+	port       uint16
+	rangeStart uint16
+	rangeEnd   uint16
+}
+
 func NewDefaultManager(fm firewall.Manager) *DefaultManager {
 	return &DefaultManager{
 		firewall:       fm,
 		peerRulesPairs: make(map[id.RuleID][]firewall.Rule),
-		routeRules:     make(map[id.RuleID]struct{}),
+		routeRules:     make(map[id.RuleID]firewall.Rule),
 	}
 }
 
@@ -88,11 +119,14 @@ func (d *DefaultManager) ApplyFiltering(networkMap *mgmProto.NetworkMap, dnsRout
 			time.Since(start), total)
 	}()
 
-	d.applyPeerACLs(networkMap)
+	peerErr := d.applyPeerACLs(networkMap)
+	if peerErr != nil {
+		log.Errorf("apply peer ACLs: %v", peerErr)
+	}
 
 	routeErr := d.applyRouteACLs(networkMap.RoutesFirewallRules, dnsRouteFeatureFlag)
 	if routeErr != nil {
-		log.Errorf("Failed to apply route ACLs: %v", routeErr)
+		log.Errorf("apply route ACLs: %v", routeErr)
 	}
 
 	flushErr := d.firewall.Flush()
@@ -104,7 +138,7 @@ func (d *DefaultManager) ApplyFiltering(networkMap *mgmProto.NetworkMap, dnsRout
 	// If applying or flushing failed, leave the previous hash untouched so the
 	// next (possibly identical) update is not skipped and gets a chance to
 	// reconcile the firewall state.
-	if err == nil && routeErr == nil && flushErr == nil {
+	if err == nil && peerErr == nil && routeErr == nil && flushErr == nil {
 		d.previousConfigHash = hash
 		d.hasAppliedConfig = true
 	} else {
@@ -116,11 +150,11 @@ func (d *DefaultManager) ApplyFiltering(networkMap *mgmProto.NetworkMap, dnsRout
 // firewall state, so an identical hash means an identical resulting ruleset.
 func (d *DefaultManager) firewallConfigHash(networkMap *mgmProto.NetworkMap, dnsRouteFeatureFlag bool) (uint64, error) {
 	return hashstructure.Hash(struct {
-		PeerRules            []*mgmProto.FirewallRule
-		PeerRulesIsEmpty     bool
-		RouteRules           []*mgmProto.RouteFirewallRule
-		RouteRulesIsEmpty    bool
-		DNSRouteFeatureFlag  bool
+		PeerRules           []*mgmProto.FirewallRule
+		PeerRulesIsEmpty    bool
+		RouteRules          []*mgmProto.RouteFirewallRule
+		RouteRulesIsEmpty   bool
+		DNSRouteFeatureFlag bool
 	}{
 		PeerRules:           networkMap.GetFirewallRules(),
 		PeerRulesIsEmpty:    networkMap.GetFirewallRulesIsEmpty(),
@@ -135,7 +169,7 @@ func (d *DefaultManager) firewallConfigHash(networkMap *mgmProto.NetworkMap, dns
 	})
 }
 
-func (d *DefaultManager) applyPeerACLs(networkMap *mgmProto.NetworkMap) {
+func (d *DefaultManager) applyPeerACLs(networkMap *mgmProto.NetworkMap) error {
 	rules := networkMap.FirewallRules
 
 	// if we got empty rules list but management not set networkMap.FirewallRulesIsEmpty flag
@@ -144,13 +178,13 @@ func (d *DefaultManager) applyPeerACLs(networkMap *mgmProto.NetworkMap) {
 		log.Warn("this peer is connected to a NetBird Management service with an older version. Allowing all traffic from connected peers")
 		rules = append(rules,
 			&mgmProto.FirewallRule{
-				PeerIP:    "0.0.0.0",
+				PeerIP:    "0.0.0.0", //nolint:staticcheck
 				Direction: mgmProto.RuleDirection_IN,
 				Action:    mgmProto.RuleAction_ACCEPT,
 				Protocol:  mgmProto.RuleProtocol_ALL,
 			},
 			&mgmProto.FirewallRule{
-				PeerIP:    "0.0.0.0",
+				PeerIP:    "0.0.0.0", //nolint:staticcheck
 				Direction: mgmProto.RuleDirection_OUT,
 				Action:    mgmProto.RuleAction_ACCEPT,
 				Protocol:  mgmProto.RuleProtocol_ALL,
@@ -158,59 +192,167 @@ func (d *DefaultManager) applyPeerACLs(networkMap *mgmProto.NetworkMap) {
 		)
 	}
 
-	newRulePairs := make(map[id.RuleID][]firewall.Rule)
-	ipsetByRuleSelectors := make(map[string]string)
+	// Group incoming single-source rules from management by their
+	// (direction, action, proto, port) selector and merge sources.
+	// One call to the firewall backend per merged rule.
+	// A deny we cannot decode would leave its traffic unblocked, so skip
+	// the whole pass and keep existing rules until the next sync.
+	groups, denyErr, err := groupPeerRules(rules)
+	if denyErr != nil {
+		return fmt.Errorf("decode deny rule sources: %w", denyErr)
+	}
 
-	// TODO: deny rules should be fatal: if a deny rule fails to apply, we must
-	// roll back all allow rules to avoid a fail-open where allowed traffic bypasses
-	// the missing deny. Currently we accumulate errors and continue.
+	newRulePairs := make(map[id.RuleID][]firewall.Rule)
 	var merr *multierror.Error
-	for _, r := range rules {
-		// if this rule is member of rule selection with more than DefaultIPsCountForSet
-		// it's IP address can be used in the ipset for firewall manager which supports it
-		selector := d.getRuleGroupingSelector(r)
-		ipsetName, ok := ipsetByRuleSelectors[selector]
-		if !ok {
-			d.ipsetCounter++
-			ipsetName = fmt.Sprintf("nb%07d", d.ipsetCounter)
-			ipsetByRuleSelectors[selector] = ipsetName
-		}
-		pairID, rulePair, err := d.protoRuleToFirewallRule(r, ipsetName)
-		if err != nil {
-			merr = multierror.Append(merr, fmt.Errorf("apply firewall rule: %w", err))
+	if err != nil {
+		merr = multierror.Append(merr, err)
+	}
+
+	// Apply denies first. A deny that fails to install is a security
+	// failure (fail-open), so if any deny errors we roll back the
+	// denies we already installed in this pass and bail out without
+	// installing any accept. Pre-existing rules stay untouched until
+	// the next successful pass clears them.
+	denies, accepts := splitDenyAccept(groups)
+	if err := d.installPeerGroups(denies, newRulePairs, true); err != nil {
+		return fmt.Errorf("install deny rules: %w", err)
+	}
+
+	if err := d.installPeerGroups(accepts, newRulePairs, false); err != nil {
+		merr = multierror.Append(merr, err)
+	}
+
+	// Tear down rules that disappeared from the networkmap. Any rule
+	// the backend refuses to delete stays in our tracking so it gets
+	// retried on the next ApplyFiltering. Otherwise a transient
+	// delete failure would leak the rule in the firewall until the
+	// process exits.
+	for pairID, rules := range d.peerRulesPairs {
+		if _, ok := newRulePairs[pairID]; ok {
 			continue
 		}
-		if len(rulePair) > 0 {
-			d.peerRulesPairs[pairID] = rulePair
-			newRulePairs[pairID] = rulePair
-		}
-	}
-
-	if merr != nil {
-		log.Errorf("failed to apply %d peer ACL rule(s): %v", merr.Len(), nberrors.FormatErrorOrNil(merr))
-	}
-
-	for pairID, rules := range d.peerRulesPairs {
-		if _, ok := newRulePairs[pairID]; !ok {
-			for _, rule := range rules {
-				if err := d.firewall.DeletePeerRule(rule); err != nil {
-					log.Errorf("failed to delete peer firewall rule: %v", err)
-					continue
-				}
+		var remaining []firewall.Rule
+		for _, rule := range rules {
+			if err := d.firewall.DeleteFilterRule(rule); err != nil {
+				merr = multierror.Append(merr, fmt.Errorf("delete peer rule, will retry: %w", err))
+				remaining = append(remaining, rule)
 			}
-			delete(d.peerRulesPairs, pairID)
+		}
+		if len(remaining) > 0 {
+			newRulePairs[pairID] = remaining
 		}
 	}
 	d.peerRulesPairs = newRulePairs
+
+	return nberrors.FormatErrorOrNil(merr)
+}
+
+// installPeerGroups applies each group and records the resulting rule
+// pairs in newRulePairs. With atomic set (deny rules), a single failure
+// rolls back every rule installed in this call and returns, leaving the
+// firewall exactly as before: denies are fail-closed and must be applied
+// all-or-nothing. With atomic unset (accept rules), failures are
+// accumulated and the remaining groups still install, so one malformed
+// allow cannot drop every other legitimate allow in the pass.
+func (d *DefaultManager) installPeerGroups(groups []*peerRuleGroup, newRulePairs map[id.RuleID][]firewall.Rule, atomic bool) error {
+	var freshlyInstalled []id.RuleID
+	var merr *multierror.Error
+	for _, g := range groups {
+		pairID, rulePair, err := d.applyPeerGroup(g)
+		if err != nil {
+			if atomic {
+				d.rollbackInstalled(freshlyInstalled)
+				return fmt.Errorf("apply firewall rule: %w", err)
+			}
+			merr = multierror.Append(merr, fmt.Errorf("apply firewall rule: %w", err))
+			continue
+		}
+		if len(rulePair) == 0 {
+			continue
+		}
+		if _, existed := d.peerRulesPairs[pairID]; !existed {
+			freshlyInstalled = append(freshlyInstalled, pairID)
+		}
+		d.peerRulesPairs[pairID] = rulePair
+		newRulePairs[pairID] = rulePair
+	}
+	return nberrors.FormatErrorOrNil(merr)
+}
+
+func (d *DefaultManager) rollbackInstalled(pairIDs []id.RuleID) {
+	var merr *multierror.Error
+	for _, pairID := range pairIDs {
+		// Keep any rule the backend refuses to delete tracked so it is
+		// retried on the next ApplyFiltering instead of leaking in the
+		// firewall with no tracking left to remove it.
+		var remaining []firewall.Rule
+		for _, rule := range d.peerRulesPairs[pairID] {
+			if err := d.firewall.DeleteFilterRule(rule); err != nil {
+				merr = multierror.Append(merr, fmt.Errorf("rule %s: %w", pairID, err))
+				remaining = append(remaining, rule)
+			}
+		}
+		if len(remaining) > 0 {
+			d.peerRulesPairs[pairID] = remaining
+		} else {
+			delete(d.peerRulesPairs, pairID)
+		}
+	}
+	if err := nberrors.FormatErrorOrNil(merr); err != nil {
+		log.Errorf("rollback peer rules: %v", err)
+	}
+}
+
+func (d *DefaultManager) applyPeerGroup(g *peerRuleGroup) (id.RuleID, []firewall.Rule, error) {
+	protocol, err := ConvertToFirewallProtocol(g.protocol)
+	if err != nil {
+		return "", nil, fmt.Errorf("skipping firewall rule: %w", err)
+	}
+	action, err := convertFirewallAction(g.action)
+	if err != nil {
+		return "", nil, fmt.Errorf("skipping firewall rule: %w", err)
+	}
+	port, err := resolveGroupPort(g)
+	if err != nil {
+		return "", nil, err
+	}
+
+	var fwRule firewall.Rule
+	switch g.direction {
+	case mgmProto.RuleDirection_IN:
+		fwRule, err = d.firewall.AddFilterRule(g.policyID, g.sources, firewall.Network{}, protocol, nil, port, action)
+	case mgmProto.RuleDirection_OUT:
+		if d.firewall.IsStateful() {
+			return "", nil, nil
+		}
+		if shouldSkipInvertedRule(protocol, port) {
+			return "", nil, nil
+		}
+		fwRule, err = d.firewall.AddFilterRule(g.policyID, g.sources, firewall.Network{}, protocol, port, nil, action)
+	default:
+		return "", nil, errors.New("invalid direction")
+	}
+
+	if err != nil {
+		return "", nil, fmt.Errorf("add firewall rule: %w", err)
+	}
+	if fwRule == nil {
+		return "", nil, fmt.Errorf("add firewall rule: %w", ErrNoRuleReturned)
+	}
+
+	// Derive the pair id from the backend rule, like the route path:
+	// the backend dedups identical content, so two policies authorizing
+	// the same flow resolve to the same id and a single backing rule.
+	return fwRule.ID(), []firewall.Rule{fwRule}, nil
 }
 
 func (d *DefaultManager) applyRouteACLs(rules []*mgmProto.RouteFirewallRule, dynamicResolver bool) error {
-	newRouteRules := make(map[id.RuleID]struct{}, len(rules))
+	newRouteRules := make(map[id.RuleID]firewall.Rule, len(rules))
 	var merr *multierror.Error
 
-	// Apply new rules - firewall manager will return existing rule ID if already present
+	// Apply new rules - firewall manager will return the existing rule if already present
 	for _, rule := range rules {
-		id, err := d.applyRouteACL(rule, dynamicResolver)
+		addedRule, err := d.applyRouteACL(rule, dynamicResolver)
 		if err != nil {
 			if errors.Is(err, ErrSourceRangesEmpty) {
 				log.Debugf("skipping empty sources rule with destination %s: %v", rule.Destination, err)
@@ -219,16 +361,18 @@ func (d *DefaultManager) applyRouteACLs(rules []*mgmProto.RouteFirewallRule, dyn
 			}
 			continue
 		}
-		newRouteRules[id] = struct{}{}
+		newRouteRules[addedRule.ID()] = addedRule
 	}
 
-	// Clean up old firewall rules
-	for id := range d.routeRules {
-		if _, exists := newRouteRules[id]; !exists {
-			if err := d.firewall.DeleteRouteRule(id); err != nil {
-				merr = multierror.Append(merr, fmt.Errorf("delete route rule: %w", err))
-			}
-			// implicitly deleted from the map
+	// Tear down old route rules; retain ones the backend refused so a
+	// transient failure doesn't leave orphaned rules in the firewall.
+	for ruleID, rule := range d.routeRules {
+		if _, exists := newRouteRules[ruleID]; exists {
+			continue
+		}
+		if err := d.firewall.DeleteFilterRule(rule); err != nil {
+			merr = multierror.Append(merr, fmt.Errorf("delete route rule, will retry: %w", err))
+			newRouteRules[ruleID] = rule
 		}
 	}
 
@@ -236,102 +380,202 @@ func (d *DefaultManager) applyRouteACLs(rules []*mgmProto.RouteFirewallRule, dyn
 	return nberrors.FormatErrorOrNil(merr)
 }
 
-func (d *DefaultManager) applyRouteACL(rule *mgmProto.RouteFirewallRule, dynamicResolver bool) (id.RuleID, error) {
+func (d *DefaultManager) applyRouteACL(rule *mgmProto.RouteFirewallRule, dynamicResolver bool) (firewall.Rule, error) {
 	if len(rule.SourceRanges) == 0 {
-		return "", ErrSourceRangesEmpty
+		return nil, ErrSourceRangesEmpty
 	}
 
 	var sources []netip.Prefix
 	for _, sourceRange := range rule.SourceRanges {
 		source, err := netip.ParsePrefix(sourceRange)
 		if err != nil {
-			return "", fmt.Errorf("parse source range: %w", err)
+			return nil, fmt.Errorf("parse source range: %w", err)
 		}
-		sources = append(sources, source)
+		sources = append(sources, firewall.UnmapPrefix(source))
 	}
 
 	destination, err := determineDestination(rule, dynamicResolver, sources)
 	if err != nil {
-		return "", fmt.Errorf("determine destination: %w", err)
+		return nil, fmt.Errorf("determine destination: %w", err)
 	}
 
-	protocol, err := convertToFirewallProtocol(rule.Protocol)
+	protocol, err := ConvertToFirewallProtocol(rule.Protocol)
 	if err != nil {
-		return "", fmt.Errorf("invalid protocol: %w", err)
+		return nil, fmt.Errorf("invalid protocol: %w", err)
 	}
 
 	action, err := convertFirewallAction(rule.Action)
 	if err != nil {
-		return "", fmt.Errorf("invalid action: %w", err)
+		return nil, fmt.Errorf("invalid action: %w", err)
 	}
 
 	dPorts := convertPortInfo(rule.PortInfo)
 
-	addedRule, err := d.firewall.AddRouteFiltering(rule.PolicyID, sources, destination, protocol, nil, dPorts, action)
+	addedRule, err := d.firewall.AddFilterRule(rule.PolicyID, sources, destination, protocol, nil, dPorts, action)
 	if err != nil {
-		return "", fmt.Errorf("add route rule: %w", err)
+		return nil, fmt.Errorf("add route rule: %w", err)
+	}
+	if addedRule == nil {
+		return nil, fmt.Errorf("add route rule: %w", ErrNoRuleReturned)
 	}
 
-	return id.RuleID(addedRule.ID()), nil
+	return addedRule, nil
 }
 
-func (d *DefaultManager) protoRuleToFirewallRule(
-	r *mgmProto.FirewallRule,
-	ipsetName string,
-) (id.RuleID, []firewall.Rule, error) {
-	ip, err := extractRuleIP(r)
-	if err != nil {
-		return "", nil, err
+// splitDenyAccept partitions groups by action so denies can be
+// applied before accepts. Order within each bucket is preserved.
+func splitDenyAccept(groups []*peerRuleGroup) (denies, accepts []*peerRuleGroup) {
+	for _, g := range groups {
+		if g.action == mgmProto.RuleAction_DROP {
+			denies = append(denies, g)
+		} else {
+			accepts = append(accepts, g)
+		}
+	}
+	return denies, accepts
+}
+
+// groupPeerRules merges single-source rules sharing a selector into
+// multi-source groups. It splits source-decode failures by action:
+// denyErr is non-nil when a deny rule could not be decoded, which is a
+// fail-open risk the caller must treat as fatal for the pass; err
+// carries the tolerable accept-rule failures the caller can log and
+// continue past.
+func groupPeerRules(rules []*mgmProto.FirewallRule) (groups []*peerRuleGroup, denyErr error, err error) {
+	var denyMerr, acceptMerr *multierror.Error
+	byKey := make(map[peerRuleKey]*peerRuleGroup)
+	order := make([]peerRuleKey, 0)
+
+	for _, r := range rules {
+		srcs, decErr := extractRuleSources(r)
+		if decErr != nil {
+			if r.Action == mgmProto.RuleAction_DROP {
+				denyMerr = multierror.Append(denyMerr, decErr)
+			} else {
+				acceptMerr = multierror.Append(acceptMerr, decErr)
+			}
+			continue
+		}
+		// A single FirewallRule normally carries one address family, but
+		// split by family defensively: each backend keys a rule to one
+		// family and would mismatch sources of the other, so a group's
+		// sources must never span families.
+		v4, v6 := splitPrefixesByFamily(srcs)
+		for _, sub := range []struct {
+			isV6    bool
+			sources []netip.Prefix
+		}{{false, v4}, {true, v6}} {
+			if len(sub.sources) == 0 {
+				continue
+			}
+			key := ruleGroupKey(r, sub.isV6)
+			g, ok := byKey[key]
+			if !ok {
+				g = &peerRuleGroup{
+					direction:  r.Direction,
+					action:     r.Action,
+					protocol:   r.Protocol,
+					port:       r.PortInfo,
+					legacyPort: r.Port,
+					policyID:   r.PolicyID,
+				}
+				byKey[key] = g
+				order = append(order, key)
+			}
+			g.sources = append(g.sources, sub.sources...)
+		}
 	}
 
-	protocol, err := convertToFirewallProtocol(r.Protocol)
-	if err != nil {
-		return "", nil, fmt.Errorf("skipping firewall rule: %s", err)
+	out := make([]*peerRuleGroup, 0, len(order))
+	for _, k := range order {
+		out = append(out, byKey[k])
+	}
+	return out, nberrors.FormatErrorOrNil(denyMerr), nberrors.FormatErrorOrNil(acceptMerr)
+}
+
+func prefixIsV6(p netip.Prefix) bool {
+	return p.Addr().Is6() && !p.Addr().Is4In6()
+}
+
+// splitPrefixesByFamily partitions prefixes into IPv4 and IPv6 groups.
+func splitPrefixesByFamily(prefixes []netip.Prefix) (v4, v6 []netip.Prefix) {
+	for _, p := range prefixes {
+		if prefixIsV6(p) {
+			v6 = append(v6, p)
+		} else {
+			v4 = append(v4, p)
+		}
+	}
+	return v4, v6
+}
+
+// ruleGroupKey builds the selector key for a rule. v6 must reflect the
+// rule's source family: mgmt emits one rule per family and mixing them
+// would break ICMP-variant selection in uspfilter.
+func ruleGroupKey(r *mgmProto.FirewallRule, v6 bool) peerRuleKey {
+	k := peerRuleKey{
+		v6:         v6,
+		policyID:   string(r.PolicyID),
+		direction:  r.Direction,
+		action:     r.Action,
+		protocol:   r.Protocol,
+		legacyPort: r.Port,
+	}
+	if pi := r.PortInfo; pi != nil {
+		k.port = uint16(pi.GetPort())
+		if rng := pi.GetRange(); rng != nil {
+			k.rangeStart = uint16(rng.GetStart())
+			k.rangeEnd = uint16(rng.GetEnd())
+		}
+	}
+	return k
+}
+
+// extractRuleSources returns all source prefixes the rule applies to.
+// New management populates sourcePrefixes; older management sets PeerIP.
+func extractRuleSources(r *mgmProto.FirewallRule) ([]netip.Prefix, error) {
+	if len(r.SourcePrefixes) > 0 {
+		out := make([]netip.Prefix, 0, len(r.SourcePrefixes))
+		for _, raw := range r.SourcePrefixes {
+			addr, err := netiputil.DecodeAddr(raw)
+			if err != nil {
+				return nil, fmt.Errorf("decode source prefix: %w", err)
+			}
+			out = append(out, netip.PrefixFrom(addr.Unmap(), addr.Unmap().BitLen()))
+		}
+		return out, nil
 	}
 
-	action, err := convertFirewallAction(r.Action)
+	peerIP := r.PeerIP //nolint:staticcheck // PeerIP is the legacy source field for old management servers
+	addr, err := netip.ParseAddr(peerIP)
 	if err != nil {
-		return "", nil, fmt.Errorf("skipping firewall rule: %s", err)
+		return nil, fmt.Errorf("parse peer IP %q: %w", peerIP, err)
 	}
+	addr = addr.Unmap()
+	// An unspecified PeerIP means "any peer" (legacy management
+	// allow-all fallback); only a /0 prefix matches any source in the
+	// backends, a full-length prefix would match nothing.
+	if addr.IsUnspecified() {
+		return []netip.Prefix{netip.PrefixFrom(addr, 0)}, nil
+	}
+	return []netip.Prefix{netip.PrefixFrom(addr, addr.BitLen())}, nil
+}
 
-	var port *firewall.Port
-	if !portInfoEmpty(r.PortInfo) {
-		port = convertPortInfo(r.PortInfo)
-	} else if r.Port != "" {
-		// old version of management, single port
-		value, err := strconv.Atoi(r.Port)
+func resolveGroupPort(g *peerRuleGroup) (*firewall.Port, error) {
+	if !portInfoEmpty(g.port) {
+		return convertPortInfo(g.port), nil
+	}
+	if g.legacyPort != "" {
+		value, err := strconv.ParseUint(g.legacyPort, 10, 16)
 		if err != nil {
-			return "", nil, fmt.Errorf("invalid port: %w", err)
+			return nil, fmt.Errorf("invalid port: %w", err)
 		}
-		port = &firewall.Port{
+		return &firewall.Port{
 			Values: []uint16{uint16(value)},
-		}
+		}, nil
 	}
-
-	ruleID := d.getPeerRuleID(ip, protocol, int(r.Direction), port, action)
-	if rulesPair, ok := d.peerRulesPairs[ruleID]; ok {
-		return ruleID, rulesPair, nil
-	}
-
-	var rules []firewall.Rule
-	switch r.Direction {
-	case mgmProto.RuleDirection_IN:
-		rules, err = d.addInRules(r.PolicyID, ip, protocol, port, action, ipsetName)
-	case mgmProto.RuleDirection_OUT:
-		if d.firewall.IsStateful() {
-			return "", nil, nil
-		}
-		// return traffic for outbound connections if firewall is stateless
-		rules, err = d.addOutRules(r.PolicyID, ip, protocol, port, action, ipsetName)
-	default:
-		return "", nil, fmt.Errorf("invalid direction, skipping firewall rule")
-	}
-
-	if err != nil {
-		return "", nil, err
-	}
-
-	return ruleID, rules, nil
+	// nolint:nilnil // a nil port legitimately means "no port restriction"
+	return nil, nil
 }
 
 func portInfoEmpty(portInfo *mgmProto.PortInfo) bool {
@@ -350,85 +594,9 @@ func portInfoEmpty(portInfo *mgmProto.PortInfo) bool {
 	}
 }
 
-func (d *DefaultManager) addInRules(
-	id []byte,
-	ip netip.Addr,
-	protocol firewall.Protocol,
-	port *firewall.Port,
-	action firewall.Action,
-	ipsetName string,
-) ([]firewall.Rule, error) {
-	rule, err := d.firewall.AddPeerFiltering(id, ip.AsSlice(), protocol, nil, port, action, ipsetName)
-	if err != nil {
-		return nil, fmt.Errorf("add firewall rule: %w", err)
-	}
-
-	return rule, nil
-}
-
-func (d *DefaultManager) addOutRules(
-	id []byte,
-	ip netip.Addr,
-	protocol firewall.Protocol,
-	port *firewall.Port,
-	action firewall.Action,
-	ipsetName string,
-) ([]firewall.Rule, error) {
-	if shouldSkipInvertedRule(protocol, port) {
-		return nil, nil
-	}
-
-	rule, err := d.firewall.AddPeerFiltering(id, ip.AsSlice(), protocol, port, nil, action, ipsetName)
-	if err != nil {
-		return nil, fmt.Errorf("add firewall rule: %w", err)
-	}
-
-	return rule, nil
-}
-
-// getPeerRuleID returns unique ID for the rule based on its parameters.
-func (d *DefaultManager) getPeerRuleID(
-	ip netip.Addr,
-	proto firewall.Protocol,
-	direction int,
-	port *firewall.Port,
-	action firewall.Action,
-) id.RuleID {
-	idStr := ip.String() + string(proto) + strconv.Itoa(direction) + strconv.Itoa(int(action))
-	if port != nil {
-		idStr += port.String()
-	}
-
-	return id.RuleID(hex.EncodeToString(md5.New().Sum([]byte(idStr))))
-}
-
-// getRuleGroupingSelector takes all rule properties except IP address to build selector
-func (d *DefaultManager) getRuleGroupingSelector(rule *mgmProto.FirewallRule) string {
-	return fmt.Sprintf("%v:%v:%v:%s:%v", strconv.Itoa(int(rule.Direction)), rule.Action, rule.Protocol, rule.Port, rule.PortInfo)
-}
-
-
-// extractRuleIP extracts the peer IP from a firewall rule.
-// If sourcePrefixes is populated (new management), decode the first entry and use its address.
-// Otherwise fall back to the deprecated PeerIP string field (old management).
-func extractRuleIP(r *mgmProto.FirewallRule) (netip.Addr, error) {
-	if len(r.SourcePrefixes) > 0 {
-		addr, err := netiputil.DecodeAddr(r.SourcePrefixes[0])
-		if err != nil {
-			return netip.Addr{}, fmt.Errorf("decode source prefix: %w", err)
-		}
-		return addr.Unmap(), nil
-	}
-
-	//nolint:staticcheck // PeerIP used for backward compatibility with old management
-	addr, err := netip.ParseAddr(r.PeerIP)
-	if err != nil {
-		return netip.Addr{}, fmt.Errorf("invalid IP address, skipping firewall rule")
-	}
-	return addr.Unmap(), nil
-}
-
-func convertToFirewallProtocol(protocol mgmProto.RuleProtocol) (firewall.Protocol, error) {
+// ConvertToFirewallProtocol maps a management rule protocol to the
+// firewall protocol type.
+func ConvertToFirewallProtocol(protocol mgmProto.RuleProtocol) (firewall.Protocol, error) {
 	switch protocol {
 	case mgmProto.RuleProtocol_TCP:
 		return firewall.ProtocolTCP, nil

client/internal/acl/manager_test.go

@@ -5,11 +5,12 @@ import (
 	"net/netip"
 	"testing"
 
-	"github.com/golang/mock/gomock"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+	"go.uber.org/mock/gomock"
 
 	"github.com/netbirdio/netbird/client/firewall"
+	fwmanager "github.com/netbirdio/netbird/client/firewall/manager"
 	"github.com/netbirdio/netbird/client/iface"
 	"github.com/netbirdio/netbird/client/iface/wgaddr"
 	"github.com/netbirdio/netbird/client/internal/acl/mocks"
@@ -77,9 +78,9 @@ func TestDefaultManager(t *testing.T) {
 	})
 
 	t.Run("add extra rules", func(t *testing.T) {
-		existedPairs := map[string]struct{}{}
+		existedPairs := map[fwmanager.RuleID]struct{}{}
 		for id := range acl.peerRulesPairs {
-			existedPairs[id.ID()] = struct{}{}
+			existedPairs[id] = struct{}{}
 		}
 
 		// remove first rule
@@ -87,7 +88,7 @@ func TestDefaultManager(t *testing.T) {
 		networkMap.FirewallRules = append(
 			networkMap.FirewallRules,
 			&mgmProto.FirewallRule{
-				PeerIP:    "10.93.0.3",
+				PeerIP:    "10.93.0.3", //nolint:staticcheck
 				Direction: mgmProto.RuleDirection_IN,
 				Action:    mgmProto.RuleAction_DROP,
 				Protocol:  mgmProto.RuleProtocol_ICMP,
@@ -106,7 +107,7 @@ func TestDefaultManager(t *testing.T) {
 		// check that old rule was removed
 		previousCount := 0
 		for id := range acl.peerRulesPairs {
-			if _, ok := existedPairs[id.ID()]; ok {
+			if _, ok := existedPairs[id]; ok {
 				previousCount++
 			}
 		}
@@ -556,12 +557,12 @@ func TestApplyFilteringSkipsUnchangedConfig(t *testing.T) {
 
 func buildNetworkMap(peerRules, routeRules int) *mgmProto.NetworkMap {
 	nm := &mgmProto.NetworkMap{
-		FirewallRulesIsEmpty:      peerRules == 0,
+		FirewallRulesIsEmpty:       peerRules == 0,
 		RoutesFirewallRulesIsEmpty: routeRules == 0,
 	}
 	for i := range peerRules {
 		nm.FirewallRules = append(nm.FirewallRules, &mgmProto.FirewallRule{
-			PeerIP:    fmt.Sprintf("10.%d.%d.%d", i>>16&0xff, i>>8&0xff, i&0xff),
+			PeerIP:    fmt.Sprintf("10.%d.%d.%d", i>>16&0xff, i>>8&0xff, i&0xff), //nolint:staticcheck
 			Direction: mgmProto.RuleDirection_IN,
 			Action:    mgmProto.RuleAction_ACCEPT,
 			Protocol:  mgmProto.RuleProtocol_TCP,

client/internal/acl/mocks/iface_mapper.go

@@ -7,7 +7,7 @@ package mocks
 import (
 	reflect "reflect"
 
-	gomock "github.com/golang/mock/gomock"
+	gomock "go.uber.org/mock/gomock"
 	wgdevice "golang.zx2c4.com/wireguard/device"
 
 	"github.com/netbirdio/netbird/client/iface/device"

client/internal/auth/auth.go

@@ -138,26 +138,37 @@ func (a *Auth) IsSSOSupported(ctx context.Context) (bool, error) {
 
 // GetOAuthFlow returns an OAuth flow (PKCE or Device) using the existing management connection
 // This avoids creating a new connection to the management server
-func (a *Auth) GetOAuthFlow(ctx context.Context, forceDeviceAuth bool) (OAuthFlow, error) {
+func (a *Auth) GetOAuthFlow(ctx context.Context, forceDeviceAuth bool, hint string) (OAuthFlow, error) {
 	var flow OAuthFlow
-	var err error
 
-	err = a.withRetry(ctx, func(client *mgm.GrpcClient) error {
+	err := a.withRetry(ctx, func(client *mgm.GrpcClient) error {
 		if forceDeviceAuth {
-			flow, err = a.getDeviceFlow(client)
-			return err
+			deviceFlow, err := a.getDeviceFlow(client)
+			if err != nil {
+				return err
+			}
+			deviceFlow.SetLoginHint(hint)
+			flow = deviceFlow
+			return nil
 		}
 
 		// Try PKCE flow first
-		flow, err = a.getPKCEFlow(client)
+		pkceFlow, err := a.getPKCEFlow(client)
 		if err != nil {
 			// If PKCE not supported, try Device flow
 			if s, ok := status.FromError(err); ok && (s.Code() == codes.NotFound || s.Code() == codes.Unimplemented) {
-				flow, err = a.getDeviceFlow(client)
-				return err
+				deviceFlow, err := a.getDeviceFlow(client)
+				if err != nil {
+					return err
+				}
+				deviceFlow.SetLoginHint(hint)
+				flow = deviceFlow
+				return nil
 			}
 			return err
 		}
+		pkceFlow.SetLoginHint(hint)
+		flow = pkceFlow
 		return nil
 	})
 
@@ -357,6 +368,7 @@ func (a *Auth) setSystemInfoFlags(info *system.Info) {
 		a.config.EnableSSHLocalPortForwarding,
 		a.config.EnableSSHRemotePortForwarding,
 		a.config.DisableSSHAuth,
+		a.config.RemoteJobsAllowed,
 	)
 }
 

client/internal/auth/device_flow.go

@@ -304,6 +304,16 @@ func (d *DeviceAuthorizationFlow) WaitToken(ctx context.Context, info AuthFlowIn
 				return TokenInfo{}, fmt.Errorf("validate access token failed with error: %v", err)
 			}
 
+			// Same as the PKCE flow: the account the token belongs to is what
+			// callers store to send back as the login_hint. Without it a client
+			// driven through the device flow — Android TV and tvOS — never binds
+			// an account to its profile and every later login goes out blind.
+			if email, err := parseEmailFromIDToken(tokenInfo.IDToken); err != nil {
+				log.Warnf("failed to parse email from ID token: %v", err)
+			} else {
+				tokenInfo.Email = email
+			}
+
 			log.Infof("device flow: user authorization confirmed after %d polls in %s", polls, time.Since(start).Round(time.Second))
 			return tokenInfo, err
 		}

client/internal/auth/oauth.go

@@ -97,9 +97,7 @@ func authenticateWithPKCEFlow(ctx context.Context, config *profilemanager.Config
 		return nil, fmt.Errorf("getting pkce authorization flow info failed with error: %v", err)
 	}
 
-	if hint != "" {
-		pkceFlowInfo.SetLoginHint(hint)
-	}
+	pkceFlowInfo.SetLoginHint(hint)
 
 	return pkceFlowInfo, nil
 }
@@ -127,9 +125,7 @@ func authenticateWithDeviceCodeFlow(ctx context.Context, config *profilemanager.
 		}
 	}
 
-	if hint != "" {
-		deviceFlowInfo.SetLoginHint(hint)
-	}
+	deviceFlowInfo.SetLoginHint(hint)
 
 	return deviceFlowInfo, nil
 }

client/internal/conn_mgr.go

@@ -2,6 +2,7 @@ package internal
 
 import (
 	"context"
+	"maps"
 	"os"
 	"strconv"
 	"sync"
@@ -14,6 +15,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/peer"
 	"github.com/netbirdio/netbird/client/internal/peerstore"
 	"github.com/netbirdio/netbird/route"
+	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
 )
 
 // lazyForce is the resolved local decision for lazy connections, layered above the
@@ -37,11 +39,13 @@ const (
 // The only exception is ActivatePeer, which is safe for concurrent use so the
 // DNS warm-up path can call it without contending on the engine mutex.
 type ConnMgr struct {
-	peerStore        *peerstore.Store
-	statusRecorder   *peer.Status
-	iface            lazyconn.WGIface
-	force            lazyForce
-	rosenpassEnabled bool
+	peerStore      *peerstore.Store
+	statusRecorder *peer.Status
+	iface          lazyconn.WGIface
+	force          lazyForce
+	// remoteLazyEnabled caches the account-wide lazy feature flag from management.
+	// It is the default for peers that do not carry a per-peer lazy hint.
+	remoteLazyEnabled bool
 
 	lazyConnMgr *manager.Manager
 	// lazyConnMgrMu guards the lazyConnMgr pointer for readers outside the
@@ -53,6 +57,10 @@ type ConnMgr struct {
 	// (re)armed (Mode A at arm time). Injected by the engine; nil disables the reconcile.
 	reconcileRoutedIPs func(peerKey string) error
 
+	// appliedExcludeList is the exclude set last handed to the lazy manager, kept so an
+	// unchanged set on the next sync skips the O(n) reconciliation.
+	appliedExcludeList map[string]bool
+
 	wg            sync.WaitGroup
 	lazyCtx       context.Context
 	lazyCtxCancel context.CancelFunc
@@ -66,78 +74,59 @@ func (e *ConnMgr) SetRoutedIPsReconciler(fn func(peerKey string) error) {
 
 func NewConnMgr(engineConfig *EngineConfig, statusRecorder *peer.Status, peerStore *peerstore.Store, iface lazyconn.WGIface) *ConnMgr {
 	e := &ConnMgr{
-		peerStore:        peerStore,
-		statusRecorder:   statusRecorder,
-		iface:            iface,
-		force:            resolveLazyForce(engineConfig.LazyConnection),
-		rosenpassEnabled: engineConfig.RosenpassEnabled,
+		peerStore:      peerStore,
+		statusRecorder: statusRecorder,
+		iface:          iface,
+		force:          resolveLazyForce(engineConfig.LazyConnection),
 	}
 	return e
 }
 
-// Start initializes the connection manager. It starts the lazy connection manager when a
-// local override forces it on; with no local override it waits for the management feature flag.
+// Start initializes the connection manager. The lazy connection manager always runs so that
+// per-peer lazy defaults (e.g. proxy peers) work even when the account flag is off; the
+// account flag and the local override decide the default lazy state per peer (see
+// PeerLazyDefault). Rosenpass peers stay lazy-capable too: their connections just never idle
+// on their own, since rosenpass rekey traffic keeps them active.
 func (e *ConnMgr) Start(ctx context.Context) {
 	if e.lazyConnMgr != nil {
 		log.Errorf("lazy connection manager is already started")
 		return
 	}
 
-	switch e.force {
-	case lazyForceOff:
-		log.Infof("lazy connection manager is disabled by local override (%s or MDM policy)", lazyconn.EnvLazyConn)
-		e.statusRecorder.UpdateLazyConnection(false)
-		return
-	case lazyForceNone:
-		log.Infof("lazy connection manager is managed by the management feature flag")
-		e.statusRecorder.UpdateLazyConnection(false)
-		return
-	}
-
-	if e.rosenpassEnabled {
-		log.Warnf("rosenpass connection manager is enabled, lazy connection manager will not be started")
-		e.statusRecorder.UpdateLazyConnection(false)
-		return
-	}
-
 	e.initLazyManager(ctx)
-	e.statusRecorder.UpdateLazyConnection(true)
+	e.statusRecorder.UpdateLazyConnection(e.PeerLazyDefault(mgmProto.LazyState_LazyStateDefault))
 }
 
-// UpdatedRemoteFeatureFlag is called when the remote feature flag is updated.
-// If enabled, it initializes the lazy connection manager and start it. Do not need to call Start() again.
-// If disabled, then it closes the lazy connection manager and open the connections to all peers.
-func (e *ConnMgr) UpdatedRemoteFeatureFlag(ctx context.Context, enabled bool) error {
-	// a local override (NB_LAZY_CONN or local config) takes precedence over management
-	if e.force != lazyForceNone {
-		return nil
+// UpdatedRemoteFeatureFlag caches the account-wide lazy feature flag. The manager itself is
+// not started or stopped here; the per-sync exclude-list reconciliation moves normal peers
+// between the lazy and always-active sets when the flag flips.
+func (e *ConnMgr) UpdatedRemoteFeatureFlag(_ context.Context, enabled bool) error {
+	e.remoteLazyEnabled = enabled
+	if e.isStartedWithLazyMgr() {
+		e.statusRecorder.UpdateLazyConnection(e.PeerLazyDefault(mgmProto.LazyState_LazyStateDefault))
+	}
+	return nil
+}
+
+// PeerLazyDefault reports whether a peer should be lazy. The local override
+// (NB_LAZY_CONN/MDM) wins over everything; without a local override the
+// management per-peer state applies (LazyStateLazy/Eager force the decision),
+// and LazyStateDefault follows the account-wide flag.
+func (e *ConnMgr) PeerLazyDefault(state mgmProto.LazyState) bool {
+	switch e.force {
+	case lazyForceOn:
+		return true
+	case lazyForceOff:
+		return false
 	}
 
-	if enabled {
-		// if the lazy connection manager is already started, do not start it again
-		if e.lazyConnMgr != nil {
-			return nil
-		}
-
-		if e.rosenpassEnabled {
-			log.Infof("rosenpass connection manager is enabled, lazy connection manager will not be started")
-			e.statusRecorder.UpdateLazyConnection(false)
-			return nil
-		}
-
-		log.Infof("lazy connection manager is enabled by the management feature flag")
-		e.initLazyManager(ctx)
-		e.statusRecorder.UpdateLazyConnection(true)
-		return e.addPeersToLazyConnManager()
-	} else {
-		if e.lazyConnMgr == nil {
-			e.statusRecorder.UpdateLazyConnection(false)
-			return nil
-		}
-		log.Infof("lazy connection manager is disabled by management feature flag")
-		e.closeManager(ctx)
-		e.statusRecorder.UpdateLazyConnection(false)
-		return nil
+	switch state {
+	case mgmProto.LazyState_LazyStateLazy:
+		return true
+	case mgmProto.LazyState_LazyStateEager:
+		return false
+	default:
+		return e.remoteLazyEnabled
 	}
 }
 
@@ -157,6 +146,13 @@ func (e *ConnMgr) SetExcludeList(ctx context.Context, peerIDs map[string]bool) {
 		return
 	}
 
+	// The exclude set is recomputed every sync but rarely changes; skip the O(n)
+	// store lookups and reconciliation when it matches what was already applied.
+	if maps.Equal(peerIDs, e.appliedExcludeList) {
+		return
+	}
+	e.appliedExcludeList = maps.Clone(peerIDs)
+
 	excludedPeers := make([]lazyconn.PeerConfig, 0, len(peerIDs))
 
 	for peerID := range peerIDs {
@@ -192,12 +188,16 @@ func (e *ConnMgr) SetExcludeList(ctx context.Context, peerIDs map[string]bool) {
 	}
 }
 
-func (e *ConnMgr) AddPeerConn(ctx context.Context, peerKey string, conn *peer.Conn) (exists bool) {
+// AddPeerConn registers a peer connection. permanent requests an always-active connection
+// (the peer belongs to the exclude set: a forwarder, or a peer that is not lazy by policy).
+// Non-permanent peers are handed to the lazy manager. The subsequent SetExcludeList call
+// reconciles membership for existing peers across flag flips.
+func (e *ConnMgr) AddPeerConn(ctx context.Context, peerKey string, conn *peer.Conn, permanent bool) (exists bool) {
 	if success := e.peerStore.AddPeerConn(peerKey, conn); !success {
 		return true
 	}
 
-	if !e.isStartedWithLazyMgr() {
+	if !e.isStartedWithLazyMgr() || permanent {
 		if err := conn.Open(ctx); err != nil {
 			conn.Log.Errorf("failed to open connection: %v", err)
 		}
@@ -296,6 +296,8 @@ func (e *ConnMgr) Close() {
 	e.lazyConnMgrMu.Lock()
 	e.lazyConnMgr = nil
 	e.lazyConnMgrMu.Unlock()
+
+	e.appliedExcludeList = nil
 }
 
 func (e *ConnMgr) initLazyManager(engineCtx context.Context) {
@@ -309,6 +311,8 @@ func (e *ConnMgr) initLazyManager(engineCtx context.Context) {
 	e.lazyCtx, e.lazyCtxCancel = context.WithCancel(engineCtx)
 	e.lazyConnMgrMu.Unlock()
 
+	e.appliedExcludeList = nil
+
 	e.wg.Add(1)
 	go func() {
 		defer e.wg.Done()
@@ -316,46 +320,6 @@ func (e *ConnMgr) initLazyManager(engineCtx context.Context) {
 	}()
 }
 
-func (e *ConnMgr) addPeersToLazyConnManager() error {
-	peers := e.peerStore.PeersPubKey()
-	lazyPeerCfgs := make([]lazyconn.PeerConfig, 0, len(peers))
-	for _, peerID := range peers {
-		var peerConn *peer.Conn
-		var exists bool
-		if peerConn, exists = e.peerStore.PeerConn(peerID); !exists {
-			log.Warnf("failed to find peer conn for peerID: %s", peerID)
-			continue
-		}
-
-		lazyPeerCfg := lazyconn.PeerConfig{
-			PublicKey:  peerID,
-			AllowedIPs: peerConn.WgConfig().AllowedIps,
-			PeerConnID: peerConn.ConnID(),
-			Log:        peerConn.Log,
-		}
-		lazyPeerCfgs = append(lazyPeerCfgs, lazyPeerCfg)
-	}
-
-	return e.lazyConnMgr.AddActivePeers(lazyPeerCfgs)
-}
-
-func (e *ConnMgr) closeManager(ctx context.Context) {
-	if e.lazyConnMgr == nil {
-		return
-	}
-
-	e.lazyCtxCancel()
-	e.wg.Wait()
-
-	e.lazyConnMgrMu.Lock()
-	e.lazyConnMgr = nil
-	e.lazyConnMgrMu.Unlock()
-
-	for _, peerID := range e.peerStore.PeersPubKey() {
-		e.peerStore.PeerConnOpen(ctx, peerID)
-	}
-}
-
 func (e *ConnMgr) isStartedWithLazyMgr() bool {
 	return e.lazyConnMgr != nil && e.lazyCtxCancel != nil
 }

client/internal/conn_mgr_test.go

@@ -16,6 +16,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/peer"
 	"github.com/netbirdio/netbird/client/internal/peerstore"
 	"github.com/netbirdio/netbird/monotime"
+	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
 )
 
 func TestResolveLazyForce(t *testing.T) {
@@ -138,4 +139,91 @@ func TestInactivityThresholdEnv(t *testing.T) {
 	}
 }
 
+func TestPeerLazyDefault(t *testing.T) {
+	tests := []struct {
+		name          string
+		force         lazyForce
+		remoteEnabled bool
+		state         mgmProto.LazyState
+		want          bool
+	}{
+		{name: "force on wins over eager state", force: lazyForceOn, state: mgmProto.LazyState_LazyStateEager, want: true},
+		{name: "force off wins over lazy state", force: lazyForceOff, remoteEnabled: true, state: mgmProto.LazyState_LazyStateLazy, want: false},
+		{name: "none, default, account off -> active", force: lazyForceNone, state: mgmProto.LazyState_LazyStateDefault, want: false},
+		{name: "none, default, account on -> lazy", force: lazyForceNone, remoteEnabled: true, state: mgmProto.LazyState_LazyStateDefault, want: true},
+		{name: "none, lazy state, account off -> lazy", force: lazyForceNone, state: mgmProto.LazyState_LazyStateLazy, want: true},
+		{name: "none, eager state, account on -> active", force: lazyForceNone, remoteEnabled: true, state: mgmProto.LazyState_LazyStateEager, want: false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			e := &ConnMgr{force: tt.force, remoteLazyEnabled: tt.remoteEnabled}
+			if got := e.PeerLazyDefault(tt.state); got != tt.want {
+				t.Fatalf("PeerLazyDefault(%v) = %v, want %v", tt.state, got, tt.want)
+			}
+		})
+	}
+}
+
 func durPtr(d time.Duration) *time.Duration { return &d }
+
+// TestToExcludedLazyPeers covers the per-peer lazy classification (proxy vs
+// normal, across the force/account-flag matrix). Forwarder-target exclusion is
+// covered by TestToExcludedLazyPeers_ForwardTarget.
+func TestToExcludedLazyPeers(t *testing.T) {
+	const (
+		normalKey = "normal"
+		lazyKey   = "lazy-state"
+		eagerKey  = "eager-state"
+	)
+
+	peers := []*mgmProto.RemotePeerConfig{
+		{WgPubKey: normalKey, AllowedIps: []string{"100.64.0.1/32"}},
+		{WgPubKey: lazyKey, AllowedIps: []string{"100.64.0.2/32"}, LazyState: mgmProto.LazyState_LazyStateLazy},
+		{WgPubKey: eagerKey, AllowedIps: []string{"100.64.0.3/32"}, LazyState: mgmProto.LazyState_LazyStateEager},
+	}
+
+	tests := []struct {
+		name          string
+		force         lazyForce
+		remoteEnabled bool
+		want          map[string]bool
+	}{
+		{
+			name:  "account off: lazy-state peer lazy, normal + eager active",
+			force: lazyForceNone, remoteEnabled: false,
+			want: map[string]bool{normalKey: true, eagerKey: true},
+		},
+		{
+			name:  "account on: only eager-state peer active",
+			force: lazyForceNone, remoteEnabled: true,
+			want: map[string]bool{eagerKey: true},
+		},
+		{
+			name:  "force off: everything active",
+			force: lazyForceOff, remoteEnabled: true,
+			want: map[string]bool{normalKey: true, lazyKey: true, eagerKey: true},
+		},
+		{
+			name:  "force on: nothing active",
+			force: lazyForceOn, remoteEnabled: false,
+			want: map[string]bool{},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			e := &Engine{connMgr: &ConnMgr{force: tt.force, remoteLazyEnabled: tt.remoteEnabled}}
+			got := e.toExcludedLazyPeers(peers)
+
+			if len(got) != len(tt.want) {
+				t.Fatalf("toExcludedLazyPeers() = %v, want %v", got, tt.want)
+			}
+			for k := range tt.want {
+				if !got[k] {
+					t.Fatalf("expected peer %s excluded, got %v", k, got)
+				}
+			}
+		})
+	}
+}

client/internal/connect.go

@@ -38,6 +38,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/updater"
 	"github.com/netbirdio/netbird/client/internal/updater/installer"
 	nbnet "github.com/netbirdio/netbird/client/net"
+	"github.com/netbirdio/netbird/client/netevents"
 	cProto "github.com/netbirdio/netbird/client/proto"
 	"github.com/netbirdio/netbird/client/ssh"
 	sshconfig "github.com/netbirdio/netbird/client/ssh/config"
@@ -70,18 +71,31 @@ type ConnectClient struct {
 	updateManager *updater.Manager
 
 	persistSyncResponse bool
+
+	// netMgr gates every reconnection loop on OS-reported network
+	// availability and sweeps connections on network change.
+	netMgr *netevents.Manager
+}
+
+// ConnectClientOption configures optional ConnectClient behavior.
+type ConnectClientOption func(*ConnectClient)
+
+// WithNetEvents injects the OS network event handling.
+func WithNetEvents(events *netevents.Manager) ConnectClientOption {
+	return func(c *ConnectClient) { c.netMgr = events }
 }
 
 func NewConnectClient(
 	ctx context.Context,
 	config *profilemanager.Config,
 	statusRecorder *peer.Status,
+	opts ...ConnectClientOption,
 ) *ConnectClient {
 	// Derive the run context here so Stop owns the cancel that unblocks the run
 	// loop. runCancel is set once at construction, so Stop can call it without
 	// racing the run loop's startup. Callers therefore need not cancel before Stop.
 	runCtx, runCancel := context.WithCancel(ctx)
-	return &ConnectClient{
+	c := &ConnectClient{
 		ctx:            runCtx,
 		runCancel:      runCancel,
 		runExited:      make(chan struct{}),
@@ -89,6 +103,10 @@ func NewConnectClient(
 		statusRecorder: statusRecorder,
 		engineMutex:    sync.Mutex{},
 	}
+	for _, opt := range opts {
+		opt(c)
+	}
+	return c
 }
 
 func (c *ConnectClient) SetUpdateManager(um *updater.Manager) {
@@ -224,7 +242,7 @@ func (c *ConnectClient) run(mobileDependency MobileDependency, runningChan chan
 	wrapErr := state.Wrap
 	myPrivateKey, err := wgtypes.ParseKey(c.config.PrivateKey)
 	if err != nil {
-		log.Errorf("failed parsing Wireguard key %s: [%s]", c.config.PrivateKey, err.Error())
+		log.Errorf("failed parsing Wireguard key: %s", err)
 		return wrapErr(err)
 	}
 
@@ -274,6 +292,13 @@ func (c *ConnectClient) run(mobileDependency MobileDependency, runningChan chan
 			return nil
 		}
 
+		// suspend connection attempts while the OS reports no usable network
+		if waited, err := c.netMgr.Wait(c.ctx); err != nil {
+			return nil
+		} else if waited {
+			backOff.Reset()
+		}
+
 		state.Set(StatusConnecting)
 
 		engineCtx, cancel := context.WithCancel(c.ctx)
@@ -285,7 +310,8 @@ func (c *ConnectClient) run(mobileDependency MobileDependency, runningChan chan
 		}()
 
 		log.Debugf("connecting to the Management service %s", c.config.ManagementURL.Host)
-		mgmClient, err := mgm.NewClient(engineCtx, c.config.ManagementURL.Host, myPrivateKey, mgmTlsEnabled)
+		mgmClient, err := mgm.NewClient(engineCtx, c.config.ManagementURL.Host, myPrivateKey, mgmTlsEnabled,
+			mgm.WithNetEvents(c.netMgr))
 		if err != nil {
 			// On daemon shutdown / Down() the parent context is cancelled
 			// and the dial fails with "context canceled". Wrapping that
@@ -360,7 +386,7 @@ func (c *ConnectClient) run(mobileDependency MobileDependency, runningChan chan
 		}()
 
 		// with the global Netbird config in hand connect (just a connection, no stream yet) Signal
-		signalClient, err := connectToSignal(engineCtx, loginResp.GetNetbirdConfig(), myPrivateKey)
+		signalClient, err := connectToSignal(engineCtx, loginResp.GetNetbirdConfig(), myPrivateKey, c.netMgr)
 		if err != nil {
 			log.Error(err)
 			return wrapErr(err)
@@ -396,7 +422,8 @@ func (c *ConnectClient) run(mobileDependency MobileDependency, runningChan chan
 			engineConfig.StateDir = filepath.Dir(path)
 		}
 
-		relayManager := relayClient.NewManager(engineCtx, relayURLs, myPrivateKey.PublicKey().String(), engineConfig.MTU)
+		relayManager := relayClient.NewManager(engineCtx, relayURLs, myPrivateKey.PublicKey().String(), engineConfig.MTU,
+			relayClient.WithNetEvents(c.netMgr))
 		c.statusRecorder.SetRelayMgr(relayManager)
 		if len(relayURLs) > 0 {
 			if token != nil {
@@ -424,6 +451,7 @@ func (c *ConnectClient) run(mobileDependency MobileDependency, runningChan chan
 			UpdateManager:  c.updateManager,
 			ClientMetrics:  c.clientMetrics,
 			MetricsCtx:     c.ctx,
+			NetMgr:         c.netMgr,
 		}, mobileDependency)
 		engine.SetSyncResponsePersistence(c.persistSyncResponse)
 		c.engine = engine
@@ -480,6 +508,16 @@ func (c *ConnectClient) run(mobileDependency MobileDependency, runningChan chan
 	// status stream stuck at Connecting.
 	err = backoff.Retry(operation, backoff.WithContext(backOff, c.ctx))
 	if err != nil {
+		// Once the client context is cancelled backoff.WithContext surfaces the
+		// bare context error, and any attempt torn down mid-flight reports the
+		// same. That cancellation is the caller asking us to stop (Stop, Down or
+		// an engine restart), so exit cleanly instead of handing back a failure
+		// the caller would have to distinguish from a real one.
+		if c.ctx.Err() != nil && errors.Is(err, context.Canceled) {
+			log.Info("exiting client retry loop, context cancelled")
+			return nil
+		}
+
 		log.Debugf("exiting client retry loop due to unrecoverable error: %s", err)
 		if s, ok := gstatus.FromError(err); ok && (s.Code() == codes.PermissionDenied) {
 			state.Set(StatusNeedsLogin)
@@ -614,6 +652,7 @@ func createEngineConfig(key wgtypes.Key, config *profilemanager.Config, peerConf
 		RosenpassEnabled:              config.RosenpassEnabled,
 		RosenpassPermissive:           config.RosenpassPermissive,
 		ServerSSHAllowed:              util.ReturnBoolWithDefaultTrue(config.ServerSSHAllowed),
+		RemoteJobsAllowed:             util.ReturnBoolWithDefaultFalse(config.RemoteJobsAllowed),
 		EnableSSHRoot:                 config.EnableSSHRoot,
 		EnableSSHSFTP:                 config.EnableSSHSFTP,
 		EnableSSHLocalPortForwarding:  config.EnableSSHLocalPortForwarding,
@@ -673,7 +712,7 @@ func selectMTU(localMTU uint16, peerMTU int32) uint16 {
 }
 
 // connectToSignal creates Signal Service client and established a connection
-func connectToSignal(ctx context.Context, wtConfig *mgmProto.NetbirdConfig, ourPrivateKey wgtypes.Key) (*signal.GrpcClient, error) {
+func connectToSignal(ctx context.Context, wtConfig *mgmProto.NetbirdConfig, ourPrivateKey wgtypes.Key, netMgr *netevents.Manager) (*signal.GrpcClient, error) {
 	var sigTLSEnabled bool
 	if wtConfig.Signal.Protocol == mgmProto.HostConfig_HTTPS {
 		sigTLSEnabled = true
@@ -681,7 +720,8 @@ func connectToSignal(ctx context.Context, wtConfig *mgmProto.NetbirdConfig, ourP
 		sigTLSEnabled = false
 	}
 
-	signalClient, err := signal.NewClient(ctx, wtConfig.Signal.Uri, ourPrivateKey, sigTLSEnabled)
+	signalClient, err := signal.NewClient(ctx, wtConfig.Signal.Uri, ourPrivateKey, sigTLSEnabled,
+		signal.WithNetEvents(netMgr))
 	if err != nil {
 		log.Errorf("error while connecting to the Signal Exchange Service %s: %s", wtConfig.Signal.Uri, err)
 		return nil, gstatus.Errorf(codes.FailedPrecondition, "failed connecting to Signal Service : %s", err)
@@ -710,6 +750,7 @@ func loginToManagement(ctx context.Context, client mgm.Client, pubSSHKey []byte,
 		config.EnableSSHLocalPortForwarding,
 		config.EnableSSHRemotePortForwarding,
 		config.DisableSSHAuth,
+		config.RemoteJobsAllowed,
 	)
 	return client.Login(sysInfo, pubSSHKey, config.DNSLabels)
 }

client/internal/connect_test.go

@@ -5,65 +5,78 @@ import (
 	"testing"
 )
 
-func Test_freePort(t *testing.T) {
-	tests := []struct {
-		name        string
-		port        int
-		want        int
-		shouldMatch bool
-	}{
-		{
-			name:        "when port is 0 use random port",
-			port:        0,
-			want:        0,
-			shouldMatch: false,
-		},
-		{
-			name:        "provided and available",
-			port:        51821,
-			want:        51821,
-			shouldMatch: true,
-		},
-		{
-			name:        "provided and not available",
-			port:        51830,
-			want:        51830,
-			shouldMatch: false,
-		},
-	}
-	c1, err := net.ListenUDP("udp", &net.UDPAddr{Port: 0})
+// probeFreePort asks the OS for a free UDP port and immediately releases it.
+// The returned number is only a hint: nothing stops another process from
+// grabbing the same port before the caller gets a chance to bind it.
+//
+// A hardcoded port number is not an option here: any fixed number can fall
+// inside the ephemeral range and be held by an unrelated process on the test
+// runner.
+func probeFreePort(t *testing.T) int {
+	t.Helper()
+
+	conn, err := net.ListenUDP("udp", &net.UDPAddr{Port: 0})
 	if err != nil {
-		t.Errorf("freePort error = %v", err)
+		t.Fatalf("failed to bind probe port: %v", err)
 	}
-	defer func(c1 *net.UDPConn) {
-		_ = c1.Close()
-	}(c1)
-
-	if tests[1].port == c1.LocalAddr().(*net.UDPAddr).Port {
-		tests[1].port++
-		tests[1].want++
-	}
-
-	tests[2].port = c1.LocalAddr().(*net.UDPAddr).Port
-	tests[2].want = c1.LocalAddr().(*net.UDPAddr).Port
-
-	for _, tt := range tests {
-
-		t.Run(tt.name, func(t *testing.T) {
-			got, err := freePort(tt.port)
-
-			if err != nil {
-				t.Errorf("got an error while getting free port: %v", err)
-			}
-
-			if tt.shouldMatch && got != tt.want {
-				t.Errorf("got a different port %v, want %v", got, tt.want)
-			}
-
-			if !tt.shouldMatch && got == tt.want {
-				t.Errorf("got the same port %v, want a different port", tt.want)
-			}
-		})
-
+	port := conn.LocalAddr().(*net.UDPAddr).Port
+	if err := conn.Close(); err != nil {
+		t.Fatalf("failed to close probe port: %v", err)
 	}
+	return port
+}
+
+func Test_freePort(t *testing.T) {
+	t.Run("when port is 0 use random port", func(t *testing.T) {
+		got, err := freePort(0)
+		if err != nil {
+			t.Fatalf("got an error while getting free port: %v", err)
+		}
+		if got == 0 {
+			t.Errorf("got port 0, want a non-zero random port")
+		}
+	})
+
+	t.Run("provided and available", func(t *testing.T) {
+		const maxAttempts = 5
+
+		// The probed port is released before freePort binds it, so an
+		// unrelated process on the test runner can grab it in between,
+		// making freePort fall back to a different port. Retry with a
+		// freshly probed port instead of failing on a lost race.
+		for attempt := 1; attempt <= maxAttempts; attempt++ {
+			candidate := probeFreePort(t)
+
+			got, err := freePort(candidate)
+			if err != nil {
+				t.Fatalf("got an error while getting free port: %v", err)
+			}
+
+			if got == candidate {
+				return
+			}
+			t.Logf("attempt %d: freePort returned %d instead of the requested %d, retrying", attempt, got, candidate)
+		}
+
+		t.Fatalf("freePort did not return the requested free port after %d attempts", maxAttempts)
+	})
+
+	t.Run("provided and not available", func(t *testing.T) {
+		busy, err := net.ListenUDP("udp", &net.UDPAddr{Port: 0})
+		if err != nil {
+			t.Fatalf("failed to bind busy port: %v", err)
+		}
+		t.Cleanup(func() {
+			_ = busy.Close()
+		})
+		busyPort := busy.LocalAddr().(*net.UDPAddr).Port
+
+		got, err := freePort(busyPort)
+		if err != nil {
+			t.Fatalf("got an error while getting free port: %v", err)
+		}
+		if got == busyPort {
+			t.Errorf("got the same port %v, want a different port", busyPort)
+		}
+	})
 }

client/internal/daemonaddr/identity.go

@@ -0,0 +1,17 @@
+package daemonaddr
+
+import "strings"
+
+// CarriesIdentity reports whether the control channel at addr conveys the
+// connecting process's identity to the daemon. A Unix socket carries peer
+// credentials and a named pipe carries the client's token. Nothing else does, TCP
+// included, and there the daemon can authorize a privileged operation for nobody
+// at all: see ResolveDaemonAddr, which says as much to anyone still reaching the
+// Windows daemon on the address it served before it had a pipe.
+//
+// A client uses this to tell whether becoming privileged would get it anywhere.
+// It answers from the scheme and nothing else, so an address it does not
+// recognise counts as carrying no identity.
+func CarriesIdentity(addr string) bool {
+	return strings.HasPrefix(addr, "unix://") || strings.HasPrefix(addr, pipeScheme)
+}

client/internal/daemonaddr/identity_test.go

@@ -0,0 +1,29 @@
+package daemonaddr
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestCarriesIdentity(t *testing.T) {
+	tests := []struct {
+		addr string
+		want bool
+	}{
+		{"unix:///var/run/netbird.sock", true},
+		{"unix:///var/run/netbird/default.sock", true},
+		{"npipe://netbird", true},
+		{`npipe://\\.\pipe\ProtectedPrefix\Administrators\netbird`, true},
+		{"tcp://127.0.0.1:41731", false},
+		{"tcp://localhost:41731", false},
+		{"", false},
+		{"/var/run/netbird.sock", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.addr, func(t *testing.T) {
+			assert.Equal(t, tt.want, CarriesIdentity(tt.addr), "address %q", tt.addr)
+		})
+	}
+}

client/internal/debug/debug.go

@@ -34,9 +34,8 @@ import (
 	"github.com/netbirdio/netbird/shared/netiputil"
 )
 
-const readmeContent = `Netbird debug bundle
-This debug bundle contains the following files.
-If the --anonymize flag is set, the files are anonymized to protect sensitive information.
+const readmeContent = `This debug bundle contains the following files.
+If anonymization is enabled (--anonymize / --anonymize-level), the files are anonymized to protect sensitive information.
 
 status.txt: Anonymized status information of the NetBird client.
 client.log: Most recent, anonymized client log file of the NetBird client.
@@ -52,6 +51,7 @@ nftables.txt: Anonymized nftables rules with packet counters across all families
 sysctls.txt: Forwarding, reverse-path filter, source-validation, and conntrack accounting sysctl values that the NetBird client may read or modify, if --system-info flag was provided (Linux only).
 resolv.conf: DNS resolver configuration from /etc/resolv.conf (Unix systems only), if --system-info flag was provided.
 scutil_dns.txt: DNS configuration from scutil --dns (macOS only), if --system-info flag was provided.
+dns_windows.txt: Anonymized NRPT rules and policy table in effect, DNS client policy, and per-interface and per-adapter DNS configuration (Windows only), if --system-info flag was provided.
 resolved_domains.txt: Anonymized resolved domain IP addresses from the status recorder.
 config.txt: Anonymized configuration information of the NetBird client.
 network_map.json: Anonymized sync response containing peer configurations, routes, DNS settings, and firewall rules.
@@ -70,21 +70,34 @@ capture.pcap: Packet capture in pcap format. Only present when capture was runni
 
 
 Anonymization Process
-The files in this bundle have been anonymized to protect sensitive information. Here's how the anonymization was applied:
+The files in this bundle have been anonymized to protect sensitive information. The level applied to this bundle is recorded at the top of this file. Here's how the anonymization was applied:
 
 IP Addresses
 
-IPv4 addresses are replaced with addresses starting from 198.51.100.0
-IPv6 addresses are replaced with addresses starting from 100::
+Default level:
+- Public IPv4 addresses are replaced with addresses starting from 198.51.100.0
+- Public IPv6 addresses are replaced with addresses starting from 2001:db8:ffff::
+- IPv6 unique local addresses (fc00::/7) are anonymized as well: their random global ID uniquely identifies the network.
+- IP addresses from internal IPv4 ranges and well-known addresses are not anonymized (e.g. 8.8.8.8, 100.64.0.0/10, addresses starting with 192.168., 172.16., 10., 169.254., fe80::).
+
+Strict level (--anonymize-level strict), in addition to the default level:
+- Private (RFC 1918), CGNAT (100.64.0.0/10), and link-local (169.254.0.0/16, fe80::/10) addresses are anonymized too.
+- Internal IPv4 addresses are replaced with addresses starting from 198.18.0.0 and internal IPv6 addresses with addresses starting from 2001:db8:1::, so internal addresses remain distinguishable from public ones.
+- Addresses are mapped in order of first appearance: subnet structure, allocation scheme, and gateway conventions are not preserved. Prefix lengths of networks are preserved.
+- Peer names in front of NetBird domains are replaced with numbered placeholders (e.g. peer-1.netbird.cloud), and subdomain labels of other domains with host-N placeholders.
+- WireGuard public keys are replaced with consistent placeholder keys.
 
-IP addresses from non public ranges and well known addresses are not anonymized (e.g. 8.8.8.8, 100.64.0.0/10, addresses starting with 192.168., 172.16., 10., etc.).
 Reoccuring IP addresses are replaced with the same anonymized address.
 
 Note: The anonymized IP addresses in the status file do not match those in the log and routes files. However, the anonymized IP addresses are consistent within the status file and across the routes and log files.
 
+MAC Addresses
+MAC addresses are replaced at every anonymization level with consistent placeholders counting up from 02:00:00:00:00:01. Broadcast, multicast, and all-zero addresses are kept. At the default level a preserved IPv6 link-local address may still embed a MAC address (EUI-64); the strict level anonymizes those addresses.
+
 Domains
 All domain names (except for the netbird domains) are replaced with randomly generated strings ending in ".domain". Anonymized domains are consistent across all files in the bundle.
 Reoccuring domain names are replaced with the same anonymized domain.
+At the strict level, the peer name labels in front of netbird domains are anonymized as well.
 
 Sync Response
 The network_map.json file contains the following anonymized information:
@@ -225,6 +238,13 @@ scutil_dns.txt (macOS only):
 - Shows DNS configuration for all network interfaces
 - Includes search domains, nameservers, and DNS resolver settings
 - All IP addresses and domain names are anonymized
+
+dns_windows.txt (Windows only):
+- Lists the NRPT rules of both policy stores, the local one and the group policy one, marking the rules the client created
+- Follows them with the policy table the resolver has loaded, which differs from the rules while a change has not been picked up yet
+- Includes the DNS client group policy, the global TCP/IP and Dnscache parameters, and the DNS values of every interface that has any
+- Ends with the resolver configuration in effect per adapter, from GetAdaptersAddresses
+- All IP addresses and domain names are anonymized
 `
 
 const (
@@ -281,6 +301,7 @@ type BundleGenerator struct {
 	cliVersion     string
 
 	anonymize         bool
+	anonymizeLevel    anonymize.Level
 	includeSystemInfo bool
 	logFileCount      uint32
 
@@ -288,7 +309,10 @@ type BundleGenerator struct {
 }
 
 type BundleConfig struct {
-	Anonymize         bool
+	Anonymize bool
+	// AnonymizeLevel selects how much the anonymizer redacts.
+	// anonymize.LevelStrict implies Anonymize.
+	AnonymizeLevel    anonymize.Level
 	IncludeSystemInfo bool
 	LogFileCount      uint32
 }
@@ -327,8 +351,11 @@ func NewBundleGenerator(deps GeneratorDependencies, cfg BundleConfig) *BundleGen
 		uiLogOpener = openLogFile
 	}
 
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+	anonymizer.SetLevel(cfg.AnonymizeLevel)
+
 	return &BundleGenerator{
-		anonymizer: anonymize.NewAnonymizer(anonymize.DefaultAddresses()),
+		anonymizer: anonymizer,
 
 		internalConfig: deps.InternalConfig,
 		statusRecorder: deps.StatusRecorder,
@@ -345,7 +372,8 @@ func NewBundleGenerator(deps GeneratorDependencies, cfg BundleConfig) *BundleGen
 		daemonVersion:  deps.DaemonVersion,
 		cliVersion:     deps.CliVersion,
 
-		anonymize:         cfg.Anonymize,
+		anonymize:         cfg.Anonymize || cfg.AnonymizeLevel >= anonymize.LevelStrict,
+		anonymizeLevel:    cfg.AnonymizeLevel,
 		includeSystemInfo: cfg.IncludeSystemInfo,
 		logFileCount:      logFileCount,
 	}
@@ -485,7 +513,13 @@ func (g *BundleGenerator) addSystemInfo() {
 }
 
 func (g *BundleGenerator) addReadme() error {
-	readmeReader := strings.NewReader(readmeContent)
+	level := "none (anonymization disabled)"
+	if g.anonymize {
+		level = g.anonymizeLevel.String()
+	}
+	header := fmt.Sprintf("Netbird debug bundle\nAnonymization level applied to this bundle: %s\n", level)
+
+	readmeReader := strings.NewReader(header + readmeContent)
 	if err := g.addFileToZip(readmeReader, "README.txt"); err != nil {
 		return fmt.Errorf("add README file to zip: %w", err)
 	}
@@ -507,9 +541,10 @@ func (g *BundleGenerator) addStatus() error {
 		fullStatus := g.statusRecorder.GetFullStatus()
 		protoFullStatus := nbstatus.ToProtoFullStatus(fullStatus)
 		overview := nbstatus.ConvertToStatusOutputOverview(protoFullStatus, nbstatus.ConvertOptions{
-			Anonymize:     g.anonymize,
-			ProfileName:   profName,
-			DaemonVersion: g.daemonVersion,
+			Anonymize:      g.anonymize,
+			AnonymizeLevel: g.anonymizeLevel,
+			ProfileName:    profName,
+			DaemonVersion:  g.daemonVersion,
 		})
 		overview.CliVersion = g.cliVersion
 		statusOutput := overview.FullDetailSummary()
@@ -662,7 +697,7 @@ func (g *BundleGenerator) addCommonConfigFields(configContent *strings.Builder)
 	configContent.WriteString("NetBird Client Configuration:\n\n")
 
 	if key, err := wgtypes.ParseKey(g.internalConfig.PrivateKey); err == nil {
-		configContent.WriteString(fmt.Sprintf("PublicKey: %s\n", key.PublicKey().String()))
+		configContent.WriteString(fmt.Sprintf("PublicKey: %s\n", g.anonymizer.AnonymizeWGKey(key.PublicKey().String())))
 	}
 	configContent.WriteString(fmt.Sprintf("WgIface: %s\n", g.internalConfig.WgIface))
 	configContent.WriteString(fmt.Sprintf("WgPort: %d\n", g.internalConfig.WgPort))
@@ -676,6 +711,9 @@ func (g *BundleGenerator) addCommonConfigFields(configContent *strings.Builder)
 	if g.internalConfig.ServerSSHAllowed != nil {
 		configContent.WriteString(fmt.Sprintf("ServerSSHAllowed: %v\n", *g.internalConfig.ServerSSHAllowed))
 	}
+	if g.internalConfig.RemoteJobsAllowed != nil {
+		configContent.WriteString(fmt.Sprintf("RemoteJobsAllowed: %v\n", *g.internalConfig.RemoteJobsAllowed))
+	}
 	if g.internalConfig.EnableSSHRoot != nil {
 		configContent.WriteString(fmt.Sprintf("EnableSSHRoot: %v\n", *g.internalConfig.EnableSSHRoot))
 	}
@@ -702,6 +740,8 @@ func (g *BundleGenerator) addCommonConfigFields(configContent *strings.Builder)
 	configContent.WriteString(fmt.Sprintf("BlockLANAccess: %v\n", g.internalConfig.BlockLANAccess))
 	configContent.WriteString(fmt.Sprintf("BlockInbound: %v\n", g.internalConfig.BlockInbound))
 	configContent.WriteString(fmt.Sprintf("DisableIPv6: %v\n", g.internalConfig.DisableIPv6))
+	configContent.WriteString(fmt.Sprintf("LocalMetricsEnabled: %v\n", g.internalConfig.LocalMetricsEnabled))
+	configContent.WriteString(fmt.Sprintf("LocalMetricsAddress: %s\n", g.internalConfig.LocalMetricsAddress))
 	configContent.WriteString(fmt.Sprintf("SyncMessageVersion: %v\n", g.internalConfig.SyncMessageVersion))
 
 	if g.internalConfig.DisableNotifications != nil {
@@ -952,6 +992,11 @@ func (g *BundleGenerator) addUpdateLogs() error {
 		}
 
 		baseName := filepath.Base(logFile)
+		data, err = g.anonymizeBytes(data)
+		if err != nil {
+			log.Warnf("skipping update log file %s: %v", baseName, err)
+			continue
+		}
 		if err := g.addFileToZip(bytes.NewReader(data), filepath.Join("update-logs", baseName)); err != nil {
 			return fmt.Errorf("add update log file %s to zip: %w", baseName, err)
 		}
@@ -979,6 +1024,13 @@ func (g *BundleGenerator) addCorruptedStateFiles() error {
 		}
 
 		fileName := filepath.Base(match)
+		// Corrupted state files usually fail structured JSON anonymization,
+		// so run them through the string anonymizer instead.
+		data, err = g.anonymizeBytes(data)
+		if err != nil {
+			log.Warnf("skipping corrupted state file %s: %v", fileName, err)
+			continue
+		}
 		if err := g.addFileToZip(bytes.NewReader(data), "corrupted_states/"+fileName); err != nil {
 			log.Warnf("Failed to add corrupted state file %s to zip: %v", fileName, err)
 			continue
@@ -990,6 +1042,27 @@ func (g *BundleGenerator) addCorruptedStateFiles() error {
 	return nil
 }
 
+// anonymizeBytes runs raw file content through the string anonymizer line by
+// line when anonymization is enabled. It errors instead of returning partial
+// content, so a caller never adds an unanonymized fallback to the bundle.
+func (g *BundleGenerator) anonymizeBytes(data []byte) ([]byte, error) {
+	if !g.anonymize {
+		return data, nil
+	}
+
+	var buf bytes.Buffer
+	scanner := bufio.NewScanner(bytes.NewReader(data))
+	scanner.Buffer(make([]byte, 1024*1024), 1024*1024)
+	for scanner.Scan() {
+		buf.WriteString(g.anonymizer.AnonymizeString(scanner.Text()))
+		buf.WriteByte('\n')
+	}
+	if err := scanner.Err(); err != nil {
+		return nil, fmt.Errorf("anonymize content: %w", err)
+	}
+	return buf.Bytes(), nil
+}
+
 func (g *BundleGenerator) addMetrics() error {
 	if g.clientMetrics == nil {
 		log.Debugf("skipping metrics in debug bundle: no metrics collector")
@@ -1462,6 +1535,7 @@ func anonymizeRemotePeer(peer *mgmProto.RemotePeerConfig, anonymizer *anonymize.
 	}
 
 	peer.Fqdn = anonymizer.AnonymizeDomain(peer.Fqdn)
+	peer.WgPubKey = anonymizer.AnonymizeWGKey(peer.WgPubKey)
 
 	anonymizeSSHConfig(peer.SshConfig)
 }

client/internal/debug/debug_linux.go

@@ -844,6 +844,10 @@ func collectSysctls() string {
 		[]string{"net.ipv4.conf.all.src_valid_mark", "net.ipv4.conf.default.src_valid_mark"},
 		listInterfaceSysctls("ipv4", "src_valid_mark")...,
 	))
+	writeSysctlGroup(&builder, "accept_ra", append(
+		[]string{"net.ipv6.conf.all.accept_ra", "net.ipv6.conf.default.accept_ra"},
+		listInterfaceSysctls("ipv6", "accept_ra")...,
+	))
 	writeSysctlGroup(&builder, "conntrack", []string{
 		"net.netfilter.nf_conntrack_acct",
 		"net.netfilter.nf_conntrack_tcp_loose",

client/internal/debug/debug_nonunix.go

@@ -1,4 +1,4 @@
-//go:build !unix
+//go:build !unix && !windows
 
 package debug
 

client/internal/debug/debug_test.go

@@ -839,12 +839,13 @@ COMMIT`
 // the excluded set with a justification.
 func TestAddConfig_AllFieldsCovered(t *testing.T) {
 	excluded := map[string]string{
-		"PrivateKey":        "sensitive: WireGuard private key",
-		"PreSharedKey":      "sensitive: WireGuard pre-shared key",
-		"SSHKey":            "sensitive: SSH private key",
-		"ClientCertKeyPair": "non-config: parsed cert pair, not serialized",
-		"Name":              "non-config: profile name is not needed for debug purposes",
-		"policy":            "non-config: in-memory MDM policy snapshot, surfaced via Config.Policy() / GetConfigResponse.MDMManagedFields",
+		"PrivateKey":           "sensitive: WireGuard private key",
+		"PreSharedKey":         "sensitive: WireGuard pre-shared key",
+		"SSHKey":               "sensitive: SSH private key",
+		"ClientCertKeyPair":    "non-config: parsed cert pair, not serialized",
+		"Name":                 "non-config: profile name is not needed for debug purposes",
+		"policy":               "non-config: in-memory MDM policy snapshot, surfaced via Config.Policy() / GetConfigResponse.MDMManagedFields",
+		"DebugBundleUploadURL": "sensitive: MDM-provided upload URL may carry credentials or query tokens; kept out of the shared bundle",
 	}
 
 	mURL, _ := url.Parse("https://api.example.com:443")
@@ -864,6 +865,7 @@ func TestAddConfig_AllFieldsCovered(t *testing.T) {
 		RosenpassEnabled:              true,
 		RosenpassPermissive:           true,
 		ServerSSHAllowed:              &bTrue,
+		RemoteJobsAllowed:             &bTrue,
 		EnableSSHRoot:                 &bTrue,
 		EnableSSHSFTP:                 &bTrue,
 		EnableSSHLocalPortForwarding:  &bTrue,
@@ -886,6 +888,7 @@ func TestAddConfig_AllFieldsCovered(t *testing.T) {
 		ClientCertPath:                "/tmp/cert",
 		ClientCertKeyPath:             "/tmp/key",
 		LazyConnection:                "on",
+		DebugBundleUploadURL:          "https://upload.example.test/bundle?token=secret",
 		MTU:                           1280,
 		DisableIPv6:                   true,
 		SyncMessageVersion:            func(v int) *int { return &v }(1),
@@ -903,6 +906,13 @@ func TestAddConfig_AllFieldsCovered(t *testing.T) {
 			g.addCommonConfigFields(&sb)
 			rendered := sb.String() + renderAddConfigSpecific(g)
 
+			// DebugBundleUploadURL is an MDM-provided value that can carry
+			// credentials or signed query tokens. It is deliberately excluded
+			// above; assert it never reaches the rendered bundle — neither the
+			// field name nor the token — in either anonymize mode.
+			assert.NotContains(t, rendered, "DebugBundleUploadURL:", "MDM upload URL field must not be serialized into the debug bundle")
+			assert.NotContains(t, rendered, "token=secret", "MDM upload URL value must not leak into the debug bundle")
+
 			val := reflect.ValueOf(cfg).Elem()
 			typ := val.Type()
 			var missing []string

client/internal/debug/debug_windows.go

@@ -0,0 +1,443 @@
+//go:build windows
+
+package debug
+
+import (
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"net/netip"
+	"strings"
+	"unsafe"
+
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/sys/windows"
+	"golang.org/x/sys/windows/registry"
+
+	nbdns "github.com/netbirdio/netbird/client/internal/dns"
+)
+
+const dnsInfoFileName = "dns_windows.txt"
+
+const (
+	gpoDNSClientRoot = `SOFTWARE\Policies\Microsoft\Windows NT\DNSClient`
+	tcpipParamsPath  = `SYSTEM\CurrentControlSet\Services\Tcpip\Parameters`
+	dnscacheParams   = `SYSTEM\CurrentControlSet\Services\Dnscache\Parameters`
+)
+
+// interfaceDNSValues are the per-interface values that decide how a name is
+// resolved and registered. Everything the DNS host manager writes is in here,
+// so a bundle shows both what we set and what it replaced.
+var interfaceDNSValues = []string{
+	"NameServer",
+	"DhcpNameServer",
+	"Domain",
+	"DhcpDomain",
+	"SearchList",
+	"RegistrationEnabled",
+	"DisableDynamicUpdate",
+	"MaxNumberOfAddressesToRegister",
+	"EnableDHCP",
+}
+
+// addDNSInfo collects and adds DNS configuration information to the archive
+func (g *BundleGenerator) addDNSInfo() error {
+	if err := g.addFileToZip(strings.NewReader(g.collectDNSInfo()), dnsInfoFileName); err != nil {
+		return fmt.Errorf("add DNS info to zip: %w", err)
+	}
+
+	return nil
+}
+
+// collectDNSInfo renders the report. Everything below it reaches the platform
+// through COM and through lazily resolved procedures, which panic when a
+// procedure is missing rather than returning an error, and a debug bundle is not
+// allowed to take the daemon down. The panic is contained here, and whatever was
+// collected before it is kept and reported with it.
+func (g *BundleGenerator) collectDNSInfo() (content string) {
+	var sb strings.Builder
+
+	defer func() {
+		if r := recover(); r != nil {
+			log.Errorf("collecting Windows DNS configuration panicked: %v", r)
+			fmt.Fprintf(&sb, "\nerror: collection stopped: %v\n", r)
+		}
+		content = sb.String()
+	}()
+
+	sb.WriteString("Windows DNS configuration\n")
+	sb.WriteString("=========================\n")
+
+	adapters, adaptersErr := adapterAddresses()
+
+	g.writeNRPTRules(&sb, "NRPT rules, local policy store", nbdns.DNSPolicyConfigRoot)
+	g.writeNRPTRules(&sb, "NRPT rules, group policy store", nbdns.GPODNSPolicyConfigRoot)
+	g.writeEffectiveNRPTPolicies(&sb)
+	g.writeRegistryKey(&sb, "DNS client group policy", gpoDNSClientRoot)
+	g.writeRegistryKey(&sb, "Global TCP/IP parameters", tcpipParamsPath)
+	g.writeRegistryKey(&sb, "Dnscache parameters", dnscacheParams)
+	g.writeInterfaceDNS(&sb, "Per-interface DNS, IPv4", nbdns.InterfaceConfigPath, adapterNames(adapters))
+	g.writeInterfaceDNS(&sb, "Per-interface DNS, IPv6", nbdns.InterfaceConfigPathV6, adapterNames(adapters))
+	g.writeAdapterDNS(&sb, adapters, adaptersErr)
+
+	return sb.String()
+}
+
+// writeNRPTRules lists every rule in a policy store, ours and any other
+// product's, since a foreign rule for the same namespace decides resolution
+// just as ours does. Rules the client wrote are marked.
+func (g *BundleGenerator) writeNRPTRules(sb *strings.Builder, title, root string) {
+	writeSection(sb, title, root)
+
+	names, err := subKeyNames(root)
+	if err != nil {
+		fmt.Fprintf(sb, "error: %v\n", err)
+		return
+	}
+
+	if len(names) == 0 {
+		sb.WriteString("no rules\n")
+		return
+	}
+
+	for _, name := range names {
+		owner := ""
+		if strings.HasPrefix(strings.ToLower(name), strings.ToLower(nbdns.NRPTKeyPrefix)) {
+			owner = "  (netbird)"
+		}
+		fmt.Fprintf(sb, "%s%s\n", name, owner)
+		g.writeValues(sb, root+`\`+name, nil, "  ")
+	}
+}
+
+// writeEffectiveNRPTPolicies reports the table the resolver answers from, which
+// the registry cannot show: a rule is written before it is loaded, and it keeps
+// being enforced after its key is gone until the resolver reloads its policy.
+func (g *BundleGenerator) writeEffectiveNRPTPolicies(sb *strings.Builder) {
+	writeSection(sb, "NRPT policy table in effect", nrptPolicyClass+"."+nrptPolicyMethod+" in "+nrptPolicyNamespace)
+
+	entries, err := effectiveNRPTPolicies()
+	if err != nil {
+		fmt.Fprintf(sb, "error: %v\n", err)
+		return
+	}
+
+	if len(entries) == 0 {
+		sb.WriteString("no policies\n")
+		return
+	}
+
+	for _, entry := range entries {
+		fmt.Fprintf(sb, "%s\n", g.anonymizeValue("Namespace", entry.namespace))
+		for _, value := range entry.values {
+			fmt.Fprintf(sb, "  %s: %s\n", value.name, g.anonymizeValue(value.name, value.value))
+		}
+	}
+}
+
+// writeInterfaceDNS reports the DNS values of every interface that has any, so
+// the netbird interface can be compared against the physical ones. The registry
+// keys the values by GUID, so each is named from the adapter list; a GUID with
+// no adapter is a leftover key of an interface that no longer exists.
+func (g *BundleGenerator) writeInterfaceDNS(sb *strings.Builder, title, root string, names map[string]string) {
+	writeSection(sb, title, root)
+
+	guids, err := subKeyNames(root)
+	if err != nil {
+		fmt.Fprintf(sb, "error: %v\n", err)
+		return
+	}
+
+	var reported int
+	for _, guid := range guids {
+		var iface strings.Builder
+		g.writeValues(&iface, root+`\`+guid, interfaceDNSValues, "  ")
+		if iface.Len() == 0 {
+			continue
+		}
+
+		name, ok := names[strings.ToLower(guid)]
+		if !ok {
+			name = "no adapter with this GUID"
+		}
+
+		reported++
+		fmt.Fprintf(sb, "%s (%s)\n%s", guid, name, iface.String())
+	}
+
+	if reported == 0 {
+		sb.WriteString("no interface holds DNS values\n")
+	}
+}
+
+// writeRegistryKey reports the values of a single key, without its subkeys.
+func (g *BundleGenerator) writeRegistryKey(sb *strings.Builder, title, path string) {
+	writeSection(sb, title, path)
+
+	var values strings.Builder
+	g.writeValues(&values, path, nil, "")
+	if values.Len() == 0 {
+		sb.WriteString("no values\n")
+		return
+	}
+
+	sb.WriteString(values.String())
+}
+
+// writeValues renders the values of a key. A nil names list reports every
+// value, otherwise only those named and present.
+func (g *BundleGenerator) writeValues(sb *strings.Builder, path string, names []string, indent string) {
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, path, registry.QUERY_VALUE)
+	switch {
+	case errors.Is(err, registry.ErrNotExist), errors.Is(err, windows.ERROR_PATH_NOT_FOUND):
+		// an absent key is the normal state for the GPO store and for
+		// interfaces without DNS settings
+		log.Debugf("HKEY_LOCAL_MACHINE\\%s does not exist", path)
+		return
+	case err != nil:
+		fmt.Fprintf(sb, "%serror: open HKEY_LOCAL_MACHINE\\%s: %v\n", indent, path, err)
+		return
+	}
+	defer closeKey(k)
+
+	if names == nil {
+		names, err = k.ReadValueNames(-1)
+		if err != nil {
+			fmt.Fprintf(sb, "%serror: read value names: %v\n", indent, err)
+			return
+		}
+	}
+
+	for _, name := range names {
+		value, err := readRegistryValue(k, name)
+		switch {
+		case errors.Is(err, registry.ErrNotExist):
+			// the caller asks for a fixed set of values, most of which a
+			// given interface does not carry
+			continue
+		case err != nil:
+			// report rather than omit: a value that is there but cannot be
+			// read reads as unset otherwise
+			fmt.Fprintf(sb, "%s%s: error: %v\n", indent, name, err)
+			continue
+		}
+
+		fmt.Fprintf(sb, "%s%s: %s\n", indent, name, g.anonymizeValue(name, value))
+	}
+}
+
+// anonymizeValue redacts a registry value according to what its name says it
+// holds. Domains and addresses are handled per entry rather than by the string
+// pass: the pass only replaces domains something else in the bundle already
+// seeded, and its address regex would eat the digit labels of a reverse zone.
+func (g *BundleGenerator) anonymizeValue(name, value string) string {
+	if !g.anonymize || value == "" {
+		return value
+	}
+
+	switch {
+	case holdsDomains(name):
+		return joinValueEntries(splitValueEntries(value), g.anonymizeDomain)
+	case holdsAddresses(name):
+		return joinValueEntries(splitValueEntries(value), g.anonymizer.AnonymizeIPString)
+	default:
+		return g.anonymizer.AnonymizeString(value)
+	}
+}
+
+// holdsDomains reports whether a value name holds domains: the domain list of
+// an NRPT rule (Name) or of the policy table (Namespace), a search list, the
+// DNS suffix values of the TCP/IP and policy keys, which all end in "Domain"
+// (Domain, DhcpDomain, NV Domain, ICSDomain), and a proxy host name.
+func holdsDomains(name string) bool {
+	lower := strings.ToLower(name)
+	return lower == "name" || lower == "namespace" || lower == "searchlist" ||
+		strings.HasSuffix(lower, "domain") || strings.HasSuffix(lower, "proxyname")
+}
+
+// holdsAddresses reports whether a value name holds DNS server addresses
+// (NameServer, DhcpNameServer, GenericDNSServers, NameServers).
+func holdsAddresses(name string) bool {
+	lower := strings.ToLower(name)
+	return strings.Contains(lower, "nameserver") || strings.Contains(lower, "dnsserver")
+}
+
+// adapterNames maps adapter GUIDs, as the registry keys the interfaces, to the
+// names an operator sees.
+func adapterNames(adapters []*windows.IpAdapterAddresses) map[string]string {
+	names := make(map[string]string, len(adapters))
+	for _, adapter := range adapters {
+		guid := windows.BytePtrToString(adapter.AdapterName)
+		names[strings.ToLower(guid)] = windows.UTF16PtrToString(adapter.FriendlyName)
+	}
+	return names
+}
+
+// writeAdapterDNS reports the resolver configuration in effect per adapter,
+// which is what the resolver uses for a name no NRPT rule matches.
+func (g *BundleGenerator) writeAdapterDNS(sb *strings.Builder, adapters []*windows.IpAdapterAddresses, err error) {
+	writeSection(sb, "Adapter DNS configuration", "GetAdaptersAddresses")
+
+	if err != nil {
+		fmt.Fprintf(sb, "error: %v\n", err)
+		return
+	}
+
+	for _, adapter := range adapters {
+		name := windows.UTF16PtrToString(adapter.FriendlyName)
+		suffix := g.anonymizeDomain(windows.UTF16PtrToString(adapter.DnsSuffix))
+
+		fmt.Fprintf(sb, "%s (index %d, oper status %d)\n", name, adapter.IfIndex, adapter.OperStatus)
+		fmt.Fprintf(sb, "  DNS suffix: %s\n", suffix)
+
+		var servers []string
+		for server := adapter.FirstDnsServerAddress; server != nil; server = server.Next {
+			addr, ok := netip.AddrFromSlice(server.Address.IP())
+			if !ok {
+				continue
+			}
+
+			addr = addr.Unmap()
+			if g.anonymize {
+				addr = g.anonymizer.AnonymizeIP(addr)
+			}
+			servers = append(servers, addr.String())
+		}
+
+		fmt.Fprintf(sb, "  DNS servers: %s\n", strings.Join(servers, ", "))
+	}
+}
+
+// anonymizeDomain anonymizes a single domain, keeping the leading dot an NRPT
+// match domain carries.
+func (g *BundleGenerator) anonymizeDomain(entry string) string {
+	if !g.anonymize {
+		return entry
+	}
+
+	domain, dot := strings.CutPrefix(entry, ".")
+	if domain == "" {
+		return entry
+	}
+
+	anonymized := g.anonymizer.AnonymizeDomain(domain)
+	if dot {
+		anonymized = "." + anonymized
+	}
+	return anonymized
+}
+
+// splitValueEntries splits a registry value that holds a list. The separator
+// differs per value: a REG_MULTI_SZ arrives joined with ", ", a SearchList is
+// comma separated and a NameServer may use commas or spaces.
+func splitValueEntries(value string) []string {
+	return strings.FieldsFunc(value, func(r rune) bool {
+		return r == ',' || r == ';' || r == ' ' || r == '\t'
+	})
+}
+
+func joinValueEntries(entries []string, anonymize func(string) string) string {
+	for i, entry := range entries {
+		entries[i] = anonymize(entry)
+	}
+	return strings.Join(entries, ", ")
+}
+
+func writeSection(sb *strings.Builder, title, source string) {
+	fmt.Fprintf(sb, "\n%s\n%s\n%s\n", title, strings.Repeat("-", len(title)), source)
+}
+
+func subKeyNames(root string) ([]string, error) {
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, root, registry.ENUMERATE_SUB_KEYS)
+	if err != nil {
+		return nil, fmt.Errorf("open HKEY_LOCAL_MACHINE\\%s: %w", root, err)
+	}
+	defer closeKey(k)
+
+	names, err := k.ReadSubKeyNames(-1)
+	if err != nil {
+		return nil, fmt.Errorf("read subkey names: %w", err)
+	}
+
+	return names, nil
+}
+
+// readRegistryValue renders a value as text regardless of its type, so an
+// unexpected type in a policy key still shows up instead of being dropped.
+func readRegistryValue(k registry.Key, name string) (string, error) {
+	_, valueType, err := k.GetValue(name, nil)
+	if err != nil {
+		return "", fmt.Errorf("get value %s: %w", name, err)
+	}
+
+	switch valueType {
+	case registry.SZ, registry.EXPAND_SZ:
+		value, _, err := k.GetStringValue(name)
+		if err != nil {
+			return "", fmt.Errorf("get string value %s: %w", name, err)
+		}
+		return value, nil
+	case registry.MULTI_SZ:
+		values, _, err := k.GetStringsValue(name)
+		if err != nil {
+			return "", fmt.Errorf("get strings value %s: %w", name, err)
+		}
+		return strings.Join(values, ", "), nil
+	case registry.DWORD, registry.QWORD:
+		value, _, err := k.GetIntegerValue(name)
+		if err != nil {
+			return "", fmt.Errorf("get integer value %s: %w", name, err)
+		}
+		return fmt.Sprintf("%d (0x%x)", value, value), nil
+	case registry.BINARY:
+		value, _, err := k.GetBinaryValue(name)
+		if err != nil {
+			return "", fmt.Errorf("get binary value %s: %w", name, err)
+		}
+		return hex.EncodeToString(value), nil
+	default:
+		return fmt.Sprintf("<unhandled registry type %d>", valueType), nil
+	}
+}
+
+// adapterAddresses returns the adapter list including DNS servers. The call
+// reports the size it needs, so grow the buffer and retry until it fits.
+func adapterAddresses() (adapters []*windows.IpAdapterAddresses, err error) {
+	// GetAdaptersAddresses is resolved on first use and panics when it is
+	// missing, so this reports it as an error and leaves the rest of the
+	// report intact.
+	defer func() {
+		if r := recover(); r != nil {
+			adapters, err = nil, fmt.Errorf("GetAdaptersAddresses: %v", r)
+		}
+	}()
+
+	const flags = windows.GAA_FLAG_SKIP_ANYCAST | windows.GAA_FLAG_SKIP_MULTICAST
+
+	size := uint32(15000)
+	for range 3 {
+		buf := make([]byte, size)
+		first := (*windows.IpAdapterAddresses)(unsafe.Pointer(&buf[0]))
+
+		err := windows.GetAdaptersAddresses(windows.AF_UNSPEC, flags, 0, first, &size)
+		if errors.Is(err, windows.ERROR_BUFFER_OVERFLOW) {
+			continue
+		}
+		if err != nil {
+			return nil, fmt.Errorf("GetAdaptersAddresses: %w", err)
+		}
+
+		for adapter := first; adapter != nil; adapter = adapter.Next {
+			adapters = append(adapters, adapter)
+		}
+		return adapters, nil
+	}
+
+	return nil, fmt.Errorf("GetAdaptersAddresses: buffer kept growing")
+}
+
+func closeKey(k registry.Key) {
+	if err := k.Close(); err != nil {
+		log.Debugf("close registry key: %v", err)
+	}
+}

client/internal/debug/debug_windows_test.go

@@ -0,0 +1,146 @@
+//go:build windows
+
+package debug
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/anonymize"
+)
+
+func newDNSValueGenerator(level anonymize.Level) *BundleGenerator {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+	anonymizer.SetLevel(level)
+
+	return &BundleGenerator{
+		anonymize:      true,
+		anonymizeLevel: level,
+		anonymizer:     anonymizer,
+	}
+}
+
+// TestAnonymizeValueByName covers the value kinds of the DNS registry keys. The
+// names decide the treatment, because the string pass alone replaces only
+// domains another part of the bundle already seeded.
+func TestAnonymizeValueByName(t *testing.T) {
+	tests := []struct {
+		name      string
+		valueName string
+		value     string
+		assert    func(t *testing.T, got string)
+	}{
+		{
+			name:      "NRPT match domains keep the leading dot",
+			valueName: "Name",
+			value:     ".internal.example.com, .corp.example.org",
+			assert: func(t *testing.T, got string) {
+				t.Helper()
+				for _, entry := range strings.Split(got, ", ") {
+					assert.True(t, strings.HasPrefix(entry, "."), "entry %q should keep its leading dot", entry)
+					assert.NotContains(t, entry, "example", "entry %q should not keep the original domain", entry)
+				}
+			},
+		},
+		{
+			name:      "any value name ending in Domain is treated as a domain",
+			valueName: "ICSDomain",
+			value:     "mshome.net",
+			assert: func(t *testing.T, got string) {
+				t.Helper()
+				assert.NotContains(t, got, "mshome", "should anonymize a domain suffix value")
+			},
+		},
+		{
+			name:      "search list is a comma separated domain list",
+			valueName: "SearchList",
+			value:     "corp.example.com,branch.example.com",
+			assert: func(t *testing.T, got string) {
+				t.Helper()
+				assert.NotContains(t, got, "example", "should anonymize every search domain")
+				assert.Len(t, strings.Split(got, ", "), 2, "should keep both search domains")
+			},
+		},
+		{
+			name:      "name servers are anonymized as addresses",
+			valueName: "DhcpNameServer",
+			value:     "203.0.113.10 8.8.8.8",
+			assert: func(t *testing.T, got string) {
+				t.Helper()
+				assert.NotContains(t, got, "203.0.113.10", "should anonymize a public resolver address")
+				// well-known resolvers stay readable at every level
+				assert.Contains(t, got, "8.8.8.8", "should keep a well-known resolver address")
+			},
+		},
+		{
+			name:      "opaque values are left to the string pass",
+			valueName: "DataBasePath",
+			value:     `%SystemRoot%\System32\drivers\etc`,
+			assert: func(t *testing.T, got string) {
+				t.Helper()
+				assert.Equal(t, `%SystemRoot%\System32\drivers\etc`, got, "should not alter a path")
+			},
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			g := newDNSValueGenerator(anonymize.LevelDefault)
+			tc.assert(t, g.anonymizeValue(tc.valueName, tc.value))
+		})
+	}
+}
+
+// TestParseNRPTPolicyTable parses the MOF text of the policy table out
+// parameters, as the provider on a client with one NRPT rule renders it.
+func TestParseNRPTPolicyTable(t *testing.T) {
+	const text = `[abstract]
+class __PARAMETERS
+{
+	[Out, EmbeddedInstance("DnsClientPolicyConfiguration"): ToSubClass, ID(2): DisableOverride ToInstance] DnsClientPolicyConfiguration cmdletOutput[] = {
+instance of DnsClientPolicyConfiguration
+{
+	DirectAccessProxyType = "NoProxy";
+	DirectAccessQueryIPsecRequired = FALSE;
+	NameEncoding = "Utf8WithoutMapping";
+	Namespace = ".0.100.in-addr.arpa";
+}, 
+instance of DnsClientPolicyConfiguration
+{
+	DirectAccessProxyType = "NoProxy";
+	NameEncoding = "Utf8WithoutMapping";
+	NameServers = {"100.0.255.254", "100.0.255.253"};
+	Namespace = ".nb.internal";
+}};
+	[in] boolean Effective;
+	[out] uint32 ReturnValue = 0;
+};
+`
+
+	entries := parseNRPTPolicyTable(text)
+	require.Len(t, entries, 2, "should parse both embedded instances")
+
+	assert.Equal(t, ".0.100.in-addr.arpa", entries[0].namespace, "should read the namespace of the first instance")
+	assert.Equal(t, ".nb.internal", entries[1].namespace, "should read the namespace of the second instance")
+
+	assert.Equal(t, []registryValue{
+		{name: "DirectAccessProxyType", value: "NoProxy"},
+		{name: "DirectAccessQueryIPsecRequired", value: "FALSE"},
+		{name: "NameEncoding", value: "Utf8WithoutMapping"},
+	}, entries[0].values, "should keep the remaining values in order")
+
+	assert.Contains(t, entries[1].values, registryValue{name: "NameServers", value: "100.0.255.254, 100.0.255.253"},
+		"should flatten a MOF array")
+
+	for _, value := range entries[1].values {
+		assert.NotContains(t, value.name, "ReturnValue", "should not read the class level parameters as values")
+	}
+}
+
+func TestParseNRPTPolicyTableEmpty(t *testing.T) {
+	assert.Empty(t, parseNRPTPolicyTable(""), "should parse no entries from empty text")
+	assert.Empty(t, parseNRPTPolicyTable("class __PARAMETERS\n{\n};\n"), "should parse no entries from a table with no instances")
+}

client/internal/debug/nrpt_windows.go

@@ -0,0 +1,317 @@
+//go:build windows
+
+package debug
+
+import (
+	"errors"
+	"fmt"
+	"runtime"
+	"strings"
+	"time"
+
+	"github.com/go-ole/go-ole"
+	"github.com/go-ole/go-ole/oleutil"
+	log "github.com/sirupsen/logrus"
+)
+
+const (
+	// The NRPT policy table is reachable through the CIM class that backs
+	// Get-DnsClientNrptPolicy. Unlike the rules in the registry, the table is
+	// what the resolver currently has loaded, which is the only way to tell an
+	// applied rule from one that is merely written, in either direction.
+	nrptPolicyNamespace = `root\Microsoft\Windows\DNS`
+	nrptPolicyClass     = "PS_DnsClientNrptPolicy"
+	nrptPolicyMethod    = "Get"
+
+	// The class has no instances, so the table comes from the out parameters
+	// of a static method call, rendered as MOF text: the embedded instances
+	// arrive as a safe array of objects, which cannot be read back through the
+	// COM bindings, and the text form carries all of them.
+	nrptPolicyInstanceKeyword = "instance of DnsClientPolicyConfiguration"
+
+	nrptPolicyTimeout = 15 * time.Second
+)
+
+// COM initialization results that leave the calling thread usable: S_FALSE for
+// a thread this process already initialized, RPC_E_CHANGED_MODE for one that
+// belongs to another apartment.
+const (
+	sFalse          = 0x00000001
+	rpcEChangedMode = 0x80010106
+)
+
+// nrptQueryInFlight admits one read of the policy table at a time. A provider
+// that stops answering keeps its goroutine and the OS thread that goroutine
+// pinned, so a later bundle reports that instead of pinning another one.
+var nrptQueryInFlight = make(chan struct{}, 1)
+
+// nrptPolicyEntry is one namespace of the effective policy table, holding the
+// values of an embedded DnsClientPolicyConfiguration instance in the order the
+// provider reported them.
+type nrptPolicyEntry struct {
+	namespace string
+	values    []registryValue
+}
+
+// registryValue is a name and its rendered value, shared by the registry and
+// policy table readers so both anonymize by value name the same way.
+type registryValue struct {
+	name  string
+	value string
+}
+
+// effectiveNRPTPolicies reads the effective NRPT table. The call is bounded
+// because a WMI provider can block indefinitely and a debug bundle must not.
+func effectiveNRPTPolicies() ([]nrptPolicyEntry, error) {
+	type result struct {
+		text string
+		err  error
+	}
+
+	select {
+	case nrptQueryInFlight <- struct{}{}:
+	default:
+		return nil, errors.New("an earlier read of the policy table has not returned")
+	}
+
+	done := make(chan result, 1)
+	go func() {
+		// the slot is released here rather than by the caller, so a read that
+		// outlives the timeout holds it until the provider answers
+		defer func() { <-nrptQueryInFlight }()
+
+		text, err := nrptPolicyTableText()
+		done <- result{text: text, err: err}
+	}()
+
+	select {
+	case res := <-done:
+		if res.err != nil {
+			return nil, res.err
+		}
+		return parseNRPTPolicyTable(res.text), nil
+	case <-time.After(nrptPolicyTimeout):
+		return nil, errors.New("read of the policy table timed out")
+	}
+}
+
+// nrptPolicyTableText calls the policy table method and returns the MOF text of
+// its out parameters.
+func nrptPolicyTableText() (text string, err error) {
+	// COM is per thread, and the collection is short lived, so the thread is
+	// pinned for the duration rather than initialized for the process.
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+
+	defer func() {
+		// The COM call chain is dynamically typed, so a provider that answers
+		// with an unexpected shape must not take the daemon down with it.
+		if r := recover(); r != nil {
+			err = fmt.Errorf("read NRPT policy table: %v", r)
+		}
+	}()
+
+	owns, err := coInitialize()
+	if err != nil {
+		return "", err
+	}
+	if owns {
+		defer ole.CoUninitialize()
+	}
+
+	locator, err := oleutil.CreateObject("WbemScripting.SWbemLocator")
+	if err != nil {
+		return "", fmt.Errorf("create WMI locator: %w", err)
+	}
+	defer locator.Release()
+
+	dispatch, err := locator.QueryInterface(ole.IID_IDispatch)
+	if err != nil {
+		return "", fmt.Errorf("query WMI locator interface: %w", err)
+	}
+	defer dispatch.Release()
+
+	service, err := dispatchCall(dispatch, "ConnectServer", nil, nrptPolicyNamespace)
+	if err != nil {
+		return "", fmt.Errorf("connect to %s: %w", nrptPolicyNamespace, err)
+	}
+	defer service.Release()
+
+	inParams, err := spawnMethodInParams(service)
+	if err != nil {
+		return "", err
+	}
+	defer inParams.Release()
+
+	// The effective table is the merge of the local and the group policy
+	// store, which is what the resolver answers from.
+	if _, err := oleutil.PutProperty(inParams, "Effective", true); err != nil {
+		return "", fmt.Errorf("set Effective parameter: %w", err)
+	}
+
+	outParams, err := dispatchCall(service, "ExecMethod", nrptPolicyClass, nrptPolicyMethod, inParams)
+	if err != nil {
+		return "", fmt.Errorf("call %s.%s: %w", nrptPolicyClass, nrptPolicyMethod, err)
+	}
+	defer outParams.Release()
+
+	textVariant, err := oleutil.CallMethod(outParams, "GetObjectText_")
+	if err != nil {
+		return "", fmt.Errorf("render policy table: %w", err)
+	}
+	defer func() {
+		if err := textVariant.Clear(); err != nil {
+			log.Debugf("clear policy table variant: %v", err)
+		}
+	}()
+
+	return textVariant.ToString(), nil
+}
+
+// spawnMethodInParams builds the in parameters instance the method needs. The
+// provider rejects the call without one, even when every parameter is optional.
+func spawnMethodInParams(service *ole.IDispatch) (*ole.IDispatch, error) {
+	class, err := dispatchCall(service, "Get", nrptPolicyClass)
+	if err != nil {
+		return nil, fmt.Errorf("get class %s: %w", nrptPolicyClass, err)
+	}
+	defer class.Release()
+
+	methods, err := dispatchProperty(class, "Methods_")
+	if err != nil {
+		return nil, fmt.Errorf("get class methods: %w", err)
+	}
+	defer methods.Release()
+
+	method, err := dispatchCall(methods, "Item", nrptPolicyMethod)
+	if err != nil {
+		return nil, fmt.Errorf("get method %s: %w", nrptPolicyMethod, err)
+	}
+	defer method.Release()
+
+	params, err := dispatchProperty(method, "InParameters")
+	if err != nil {
+		return nil, fmt.Errorf("get method parameters: %w", err)
+	}
+	defer params.Release()
+
+	inParams, err := dispatchCall(params, "SpawnInstance_")
+	if err != nil {
+		return nil, fmt.Errorf("spawn parameter instance: %w", err)
+	}
+
+	return inParams, nil
+}
+
+// parseNRPTPolicyTable pulls the embedded instances out of the MOF text. Each
+// instance is a namespace of the table, with one name and value per line.
+func parseNRPTPolicyTable(text string) []nrptPolicyEntry {
+	var entries []nrptPolicyEntry
+	var current *nrptPolicyEntry
+
+	for _, line := range strings.Split(text, "\n") {
+		line = strings.TrimSpace(strings.TrimSuffix(strings.TrimSpace(line), ";"))
+
+		switch {
+		case strings.HasPrefix(line, nrptPolicyInstanceKeyword):
+			entries = append(entries, nrptPolicyEntry{})
+			current = &entries[len(entries)-1]
+			continue
+		case strings.HasPrefix(line, "}"):
+			// closes an instance, and the array with the last one, so the
+			// class level parameters that follow are not read as values
+			current = nil
+			continue
+		case current == nil, line == "{":
+			continue
+		}
+
+		name, value, ok := strings.Cut(line, " = ")
+		if !ok {
+			continue
+		}
+
+		value = unquoteMOFValue(value)
+		if name == "Namespace" {
+			current.namespace = value
+			continue
+		}
+
+		current.values = append(current.values, registryValue{name: name, value: value})
+	}
+
+	return entries
+}
+
+// unquoteMOFValue renders a MOF scalar or array as plain text: "a" becomes a,
+// and {"a", "b"} becomes a, b.
+func unquoteMOFValue(value string) string {
+	value = strings.TrimSpace(value)
+
+	if inner, ok := strings.CutPrefix(value, "{"); ok {
+		value = strings.TrimSuffix(inner, "}")
+
+		entries := strings.Split(value, ",")
+		for i, entry := range entries {
+			entries[i] = strings.Trim(strings.TrimSpace(entry), `"`)
+		}
+		return strings.Join(entries, ", ")
+	}
+
+	return strings.Trim(value, `"`)
+}
+
+// coInitialize prepares the calling thread for COM and reports whether this
+// call owns the initialization, which decides whether it may be balanced with
+// CoUninitialize. S_FALSE took a reference on a thread this process had already
+// initialized and so has to be released, while RPC_E_CHANGED_MODE took none:
+// the thread belongs to another apartment, which is usable but is not ours to
+// uninitialize.
+func coInitialize() (bool, error) {
+	err := ole.CoInitializeEx(0, ole.COINIT_MULTITHREADED)
+	if err == nil {
+		return true, nil
+	}
+
+	var oleErr *ole.OleError
+	if errors.As(err, &oleErr) {
+		switch oleErr.Code() {
+		case sFalse:
+			return true, nil
+		case rpcEChangedMode:
+			return false, nil
+		}
+	}
+
+	return false, fmt.Errorf("initialize COM: %w", err)
+}
+
+// dispatchCall calls a COM method that returns an object.
+func dispatchCall(dispatch *ole.IDispatch, method string, params ...any) (*ole.IDispatch, error) {
+	variant, err := oleutil.CallMethod(dispatch, method, params...)
+	if err != nil {
+		return nil, err
+	}
+
+	object := variant.ToIDispatch()
+	if object == nil {
+		return nil, fmt.Errorf("%s returned no object", method)
+	}
+
+	return object, nil
+}
+
+// dispatchProperty reads a COM property that holds an object.
+func dispatchProperty(dispatch *ole.IDispatch, property string) (*ole.IDispatch, error) {
+	variant, err := oleutil.GetProperty(dispatch, property)
+	if err != nil {
+		return nil, err
+	}
+
+	object := variant.ToIDispatch()
+	if object == nil {
+		return nil, fmt.Errorf("property %s holds no object", property)
+	}
+
+	return object, nil
+}

client/internal/debug/wgshow.go

@@ -35,14 +35,14 @@ func (g *BundleGenerator) toWGShowFormat(s *configurer.Stats) string {
 	var sb strings.Builder
 
 	sb.WriteString(fmt.Sprintf("interface: %s\n", s.DeviceName))
-	sb.WriteString(fmt.Sprintf("  public key: %s\n", s.PublicKey))
+	sb.WriteString(fmt.Sprintf("  public key: %s\n", g.anonymizer.AnonymizeWGKey(s.PublicKey)))
 	sb.WriteString(fmt.Sprintf("  listen port: %d\n", s.ListenPort))
 	if s.FWMark != 0 {
 		sb.WriteString(fmt.Sprintf("  fwmark: %#x\n", s.FWMark))
 	}
 
 	for _, peer := range s.Peers {
-		sb.WriteString(fmt.Sprintf("\npeer: %s\n", peer.PublicKey))
+		sb.WriteString(fmt.Sprintf("\npeer: %s\n", g.anonymizer.AnonymizeWGKey(peer.PublicKey)))
 		if peer.Endpoint.IP != nil {
 			if g.anonymize {
 				anonEndpoint := g.anonymizer.AnonymizeUDPAddr(peer.Endpoint)
@@ -54,7 +54,11 @@ func (g *BundleGenerator) toWGShowFormat(s *configurer.Stats) string {
 		if len(peer.AllowedIPs) > 0 {
 			var ipStrings []string
 			for _, ipnet := range peer.AllowedIPs {
-				ipStrings = append(ipStrings, ipnet.String())
+				ipStr := ipnet.String()
+				if g.anonymize {
+					ipStr = g.anonymizer.AnonymizeIPString(ipStr)
+				}
+				ipStrings = append(ipStrings, ipStr)
 			}
 			sb.WriteString(fmt.Sprintf("  allowed ips: %s\n", strings.Join(ipStrings, ", ")))
 		}

client/internal/dns/host_darwin.go

@@ -267,18 +267,38 @@ func (s *systemConfigurator) getSystemDNSSettings() (SystemDNSSettings, error) {
 		return SystemDNSSettings{}, fmt.Errorf("sending the command: %w", err)
 	}
 
-	var dnsSettings SystemDNSSettings
+	dnsSettings, serverAddresses, err := parseSystemDNSSettings(b)
+	if err != nil {
+		return dnsSettings, err
+	}
+
+	s.mu.Lock()
+	s.origNameservers = serverAddresses
+	s.mu.Unlock()
+
+	return dnsSettings, nil
+}
+
+// parseSystemDNSSettings parses the output of `scutil show State:/Network/Service/<id>/DNS`.
+// Lines that don't match the expected "index : value" shape are skipped: hosts with unusual
+// network services (e.g. orphaned hardware ports) can produce entries without a value.
+func parseSystemDNSSettings(out []byte) (SystemDNSSettings, []netip.Addr, error) {
+	// port is not exposed by scutil, default to 53
+	dnsSettings := SystemDNSSettings{ServerPort: DefaultPort}
 	var serverAddresses []netip.Addr
 	inSearchDomainsArray := false
 	inServerAddressesArray := false
 
-	scanner := bufio.NewScanner(bytes.NewReader(b))
+	scanner := bufio.NewScanner(bytes.NewReader(out))
 	for scanner.Scan() {
 		line := strings.TrimSpace(scanner.Text())
 		switch {
 		case strings.HasPrefix(line, "DomainName :"):
-			domainName := strings.TrimSpace(strings.Split(line, ":")[1])
-			dnsSettings.Domains = append(dnsSettings.Domains, domainName)
+			domainName := strings.TrimSpace(strings.TrimPrefix(line, "DomainName :"))
+			if domainName != "" {
+				dnsSettings.Domains = append(dnsSettings.Domains, domainName)
+			}
+			continue
 		case line == "SearchDomains : <array> {":
 			inSearchDomainsArray = true
 			continue
@@ -288,36 +308,45 @@ func (s *systemConfigurator) getSystemDNSSettings() (SystemDNSSettings, error) {
 		case line == "}":
 			inSearchDomainsArray = false
 			inServerAddressesArray = false
+			continue
+		}
+
+		if !inSearchDomainsArray && !inServerAddressesArray {
+			continue
+		}
+
+		parts := strings.SplitN(line, " : ", 2)
+		if len(parts) != 2 {
+			log.Debugf("skipping unexpected scutil DNS line %q", line)
+			continue
+		}
+		value := strings.TrimSpace(parts[1])
+		if value == "" {
+			continue
 		}
 
 		if inSearchDomainsArray {
-			searchDomain := strings.Split(line, " : ")[1]
-			dnsSettings.Domains = append(dnsSettings.Domains, searchDomain)
-		} else if inServerAddressesArray {
-			address := strings.Split(line, " : ")[1]
-			if ip, err := netip.ParseAddr(address); err == nil && !ip.IsUnspecified() {
-				ip = ip.Unmap()
-				serverAddresses = append(serverAddresses, ip)
-				// Prefer the first IPv4 server as ServerIP since our DNS listener is IPv4.
-				if !dnsSettings.ServerIP.IsValid() && ip.Is4() {
-					dnsSettings.ServerIP = ip
-				}
-			}
+			dnsSettings.Domains = append(dnsSettings.Domains, value)
+			continue
+		}
+
+		ip, err := netip.ParseAddr(value)
+		if err != nil || ip.IsUnspecified() {
+			continue
+		}
+		ip = ip.Unmap()
+		serverAddresses = append(serverAddresses, ip)
+		// Prefer the first IPv4 server as ServerIP since our DNS listener is IPv4.
+		if !dnsSettings.ServerIP.IsValid() && ip.Is4() {
+			dnsSettings.ServerIP = ip
 		}
 	}
 
 	if err := scanner.Err(); err != nil {
-		return dnsSettings, err
+		return dnsSettings, serverAddresses, err
 	}
 
-	// default to 53 port
-	dnsSettings.ServerPort = DefaultPort
-
-	s.mu.Lock()
-	s.origNameservers = serverAddresses
-	s.mu.Unlock()
-
-	return dnsSettings, nil
+	return dnsSettings, serverAddresses, nil
 }
 
 func (s *systemConfigurator) getOriginalNameservers() []netip.Addr {
@@ -435,11 +464,15 @@ func (s *systemConfigurator) getPrimaryService() (string, string, error) {
 	router := ""
 	for scanner.Scan() {
 		text := scanner.Text()
+		parts := strings.SplitN(text, ":", 2)
+		if len(parts) != 2 {
+			continue
+		}
 		if strings.Contains(text, "PrimaryService") {
-			primaryService = strings.TrimSpace(strings.Split(text, ":")[1])
+			primaryService = strings.TrimSpace(parts[1])
 		}
 		if strings.Contains(text, "Router") {
-			router = strings.TrimSpace(strings.Split(text, ":")[1])
+			router = strings.TrimSpace(parts[1])
 		}
 	}
 	if err := scanner.Err(); err != nil && err != io.EOF {

client/internal/dns/host_darwin_test.go

@@ -328,6 +328,120 @@ func removeTestDNSKey(key string) error {
 	return err
 }
 
+func TestParseSystemDNSSettings(t *testing.T) {
+	tests := []struct {
+		name            string
+		output          string
+		expectedDomains []string
+		expectedServers []netip.Addr
+		expectedIP      netip.Addr
+	}{
+		{
+			name: "well_formed",
+			output: `<dictionary> {
+  DomainName : example.com
+  SearchDomains : <array> {
+    0 : example.com
+    1 : corp.example.com
+  }
+  ServerAddresses : <array> {
+    0 : 192.168.1.1
+    1 : fd00::53
+  }
+}
+`,
+			expectedDomains: []string{"example.com", "example.com", "corp.example.com"},
+			expectedServers: []netip.Addr{netip.MustParseAddr("192.168.1.1"), netip.MustParseAddr("fd00::53")},
+			expectedIP:      netip.MustParseAddr("192.168.1.1"),
+		},
+		{
+			// entries without a value after the separator used to panic with
+			// "index out of range [1] with length 1"
+			name: "malformed_array_entries_skipped",
+			output: `<dictionary> {
+  SearchDomains : <array> {
+    0 :
+    (null)
+
+    1 : corp.example.com
+  }
+  ServerAddresses : <array> {
+    0 :
+    1 : 192.168.1.1
+  }
+}
+`,
+			expectedDomains: []string{"corp.example.com"},
+			expectedServers: []netip.Addr{netip.MustParseAddr("192.168.1.1")},
+			expectedIP:      netip.MustParseAddr("192.168.1.1"),
+		},
+		{
+			name: "domain_name_without_value_skipped",
+			output: `<dictionary> {
+  DomainName :
+  ServerAddresses : <array> {
+    0 : 192.168.1.1
+  }
+}
+`,
+			expectedServers: []netip.Addr{netip.MustParseAddr("192.168.1.1")},
+			expectedIP:      netip.MustParseAddr("192.168.1.1"),
+		},
+		{
+			name: "ipv6_first_prefers_ipv4_server_ip",
+			output: `<dictionary> {
+  ServerAddresses : <array> {
+    0 : fd00::53
+    1 : 192.168.1.1
+  }
+}
+`,
+			expectedServers: []netip.Addr{netip.MustParseAddr("fd00::53"), netip.MustParseAddr("192.168.1.1")},
+			expectedIP:      netip.MustParseAddr("192.168.1.1"),
+		},
+		{
+			name: "invalid_and_unspecified_addresses_skipped",
+			output: `<dictionary> {
+  ServerAddresses : <array> {
+    0 : (null)
+    1 : 0.0.0.0
+    2 : 192.168.1.1
+  }
+}
+`,
+			expectedServers: []netip.Addr{netip.MustParseAddr("192.168.1.1")},
+			expectedIP:      netip.MustParseAddr("192.168.1.1"),
+		},
+		{
+			name: "v4_mapped_address_unmapped",
+			output: `<dictionary> {
+  ServerAddresses : <array> {
+    0 : ::ffff:192.168.1.1
+  }
+}
+`,
+			expectedServers: []netip.Addr{netip.MustParseAddr("192.168.1.1")},
+			expectedIP:      netip.MustParseAddr("192.168.1.1"),
+		},
+		{
+			name:   "empty_output",
+			output: "",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			settings, servers, err := parseSystemDNSSettings([]byte(tc.output))
+			require.NoError(t, err, "parsing should not fail")
+
+			assert.Equal(t, tc.expectedDomains, settings.Domains, "domains should match")
+			assert.Equal(t, tc.expectedServers, servers, "server addresses should match")
+			assert.Equal(t, tc.expectedIP, settings.ServerIP, "server IP should match")
+			assert.Equal(t, DefaultPort, settings.ServerPort, "server port should default to 53")
+		})
+	}
+}
+
 func TestGetOriginalNameservers(t *testing.T) {
 	configurator := &systemConfigurator{
 		createdKeys: make(map[string]struct{}),

client/internal/dns/host_windows.go

@@ -6,8 +6,10 @@ import (
 	"fmt"
 	"io"
 	"net/netip"
+	"os"
 	"os/exec"
 	"slices"
+	"strconv"
 	"strings"
 	"syscall"
 	"time"
@@ -31,10 +33,52 @@ var (
 	dnsFlushResolverCacheFn = dnsapi.NewProc("DnsFlushResolverCache")
 )
 
+// Registry locations of the host DNS configuration this package programs,
+// exported so a diagnostic reader reports the same locations that are written.
 const (
-	dnsPolicyConfigMatchPath    = `SYSTEM\CurrentControlSet\Services\Dnscache\Parameters\DnsPolicyConfig\NetBird-Match`
-	gpoDnsPolicyRoot            = `SOFTWARE\Policies\Microsoft\Windows NT\DNSClient\DnsPolicyConfig`
-	gpoDnsPolicyConfigMatchPath = gpoDnsPolicyRoot + `\NetBird-Match`
+	// NRPTKeyPrefix starts the name of every NRPT rule key this client creates:
+	// the match rules, the catch-all, and the .local exemption. Cleanup
+	// enumerates by this prefix, so a new kind of rule is removed by existing
+	// code as long as its key starts here.
+	NRPTKeyPrefix = "NetBird-"
+
+	// nrptMatchKeyName names the match-domain rules. Older versions used
+	// different layouts under the same name: a single unsuffixed key, then one
+	// key per domain, now one key per batch of domains.
+	nrptMatchKeyName = NRPTKeyPrefix + "Match"
+
+	// DNSPolicyConfigRoot holds the NRPT rules of the local policy store.
+	DNSPolicyConfigRoot = `SYSTEM\CurrentControlSet\Services\Dnscache\Parameters\DnsPolicyConfig`
+
+	// GPODNSPolicyConfigRoot holds the NRPT rules of the group policy store,
+	// which takes precedence over the local one when it is present.
+	GPODNSPolicyConfigRoot = `SOFTWARE\Policies\Microsoft\Windows NT\DNSClient\DnsPolicyConfig`
+
+	// InterfaceConfigPath and InterfaceConfigPathV6 hold the per-interface DNS
+	// settings, keyed by interface GUID, in separate hives per address family.
+	InterfaceConfigPath   = `SYSTEM\CurrentControlSet\Services\Tcpip\Parameters\Interfaces`
+	InterfaceConfigPathV6 = `SYSTEM\CurrentControlSet\Services\Tcpip6\Parameters\Interfaces`
+)
+
+const (
+	dnsPolicyConfigMatchPath    = DNSPolicyConfigRoot + `\` + nrptMatchKeyName
+	gpoDnsPolicyConfigMatchPath = GPODNSPolicyConfigRoot + `\` + nrptMatchKeyName
+
+	dnsPolicyConfigExemptLocalPath    = DNSPolicyConfigRoot + `\` + NRPTKeyPrefix + `ExemptLocal`
+	gpoDnsPolicyConfigExemptLocalPath = GPODNSPolicyConfigRoot + `\` + NRPTKeyPrefix + `ExemptLocal`
+
+	nrptCatchAllNamespace = "."
+	// nrptLocalNamespace is reserved for multicast DNS by RFC 6762: a unicast
+	// resolver must not answer for it. The catch-all rule would hand it to us
+	// anyway, so it gets an exemption rule of its own.
+	nrptLocalNamespace = ".local"
+
+	// envLegacyDNSResolution restores the pre-catch-all behaviour: the adapter's
+	// NameServer alone, leaving the OS free to query other adapters' resolvers in
+	// parallel. An escape hatch for setups that depend on a resolver of theirs
+	// still being reachable while connected, at the cost of the leak and of the
+	// race the catch-all rule exists to close.
+	envLegacyDNSResolution = "NB_USE_LEGACY_DNS_RESOLUTION"
 
 	dnsPolicyConfigVersionKey           = "Version"
 	dnsPolicyConfigVersionValue         = 2
@@ -45,8 +89,6 @@ const (
 
 	nrptMaxDomainsPerRule = 50
 
-	interfaceConfigPath           = `SYSTEM\CurrentControlSet\Services\Tcpip\Parameters\Interfaces`
-	interfaceConfigPathV6         = `SYSTEM\CurrentControlSet\Services\Tcpip6\Parameters\Interfaces`
 	interfaceConfigNameServerKey  = "NameServer"
 	interfaceConfigDhcpNameSrvKey = "DhcpNameServer"
 	interfaceConfigSearchListKey  = "SearchList"
@@ -73,7 +115,6 @@ type registryConfigurator struct {
 	guid            string
 	routingAll      bool
 	gpo             bool
-	nrptEntryCount  int
 	origNameservers []netip.Addr
 }
 
@@ -84,7 +125,7 @@ func newHostManager(wgInterface WGIface) (*registryConfigurator, error) {
 	}
 
 	var useGPO bool
-	k, err := registry.OpenKey(registry.LOCAL_MACHINE, gpoDnsPolicyRoot, registry.QUERY_VALUE)
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, GPODNSPolicyConfigRoot, registry.QUERY_VALUE)
 	if err != nil {
 		log.Debugf("failed to open GPO DNS policy root: %v", err)
 	} else {
@@ -123,7 +164,7 @@ func (r *registryConfigurator) captureOriginalNameservers() ([]netip.Addr, error
 	seen := make(map[netip.Addr]struct{})
 	var out []netip.Addr
 	var merr *multierror.Error
-	for _, root := range []string{interfaceConfigPath, interfaceConfigPathV6} {
+	for _, root := range []string{InterfaceConfigPath, InterfaceConfigPathV6} {
 		addrs, err := r.captureFromTcpipRoot(root)
 		if err != nil {
 			merr = multierror.Append(merr, fmt.Errorf("%s: %w", root, err))
@@ -276,6 +317,13 @@ func (r *registryConfigurator) disableWINSForInterface() error {
 }
 
 func (r *registryConfigurator) applyDNSConfig(config HostDNSConfig, stateManager *statemanager.Manager) error {
+	// Clear every rule the previous apply installed before installing any new
+	// one, including a leftover catch-all: removal is unconditional so a rule
+	// from an earlier run cannot survive into a config that no longer wants it.
+	if err := r.removeDNSMatchPolicies(); err != nil {
+		log.Errorf("cleanup old dns match policies: %s", err)
+	}
+
 	if config.RouteAll {
 		if err := r.addDNSSetupForAll(config.ServerIP); err != nil {
 			return fmt.Errorf("add dns setup: %w", err)
@@ -301,19 +349,28 @@ func (r *registryConfigurator) applyDNSConfig(config HostDNSConfig, stateManager
 		matchDomains = append(matchDomains, "."+strings.TrimSuffix(dConf.Domain, "."))
 	}
 
-	if err := r.removeDNSMatchPolicies(); err != nil {
-		log.Errorf("cleanup old dns match policies: %s", err)
+	// The root namespace is a match domain like any other: it just happens to
+	// match every name. Without it the adapter's NameServer only adds one more
+	// resolver to the set Windows queries in parallel, keeping whichever answer
+	// comes back first — which leaks every query to the local network and lets a
+	// resolver other than ours answer for a name we are authoritative for.
+	if config.RouteAll {
+		if parseBoolEnv(envLegacyDNSResolution) {
+			log.Infof("%s is set, leaving DNS resolution shared with the other adapters' resolvers instead of forcing it through %s", envLegacyDNSResolution, config.ServerIP)
+		} else {
+			matchDomains = append(matchDomains, nrptCatchAllNamespace)
+			log.Infof("routing every namespace through %s: DNS resolution is now exclusive to NetBird", config.ServerIP)
+
+			if err := r.addDNSExemptLocalPolicy(); err != nil {
+				return fmt.Errorf("add dns exempt policy: %w", err)
+			}
+		}
 	}
 
 	if len(matchDomains) != 0 {
-		count, err := r.addDNSMatchPolicy(matchDomains, config.ServerIP)
-		// Update count even on error to ensure cleanup covers partially created rules
-		r.nrptEntryCount = count
-		if err != nil {
+		if err := r.addDNSMatchPolicy(matchDomains, config.ServerIP); err != nil {
 			return fmt.Errorf("add dns match policy: %w", err)
 		}
-	} else {
-		r.nrptEntryCount = 0
 	}
 
 	r.updateState(stateManager)
@@ -329,9 +386,8 @@ func (r *registryConfigurator) applyDNSConfig(config HostDNSConfig, stateManager
 
 func (r *registryConfigurator) updateState(stateManager *statemanager.Manager) {
 	if err := stateManager.UpdateState(&ShutdownState{
-		Guid:           r.guid,
-		GPO:            r.gpo,
-		NRPTEntryCount: r.nrptEntryCount,
+		Guid: r.guid,
+		GPO:  r.gpo,
 	}); err != nil {
 		log.Errorf("failed to update shutdown state: %s", err)
 	}
@@ -346,7 +402,7 @@ func (r *registryConfigurator) addDNSSetupForAll(ip netip.Addr) error {
 	return nil
 }
 
-func (r *registryConfigurator) addDNSMatchPolicy(domains []string, ip netip.Addr) (int, error) {
+func (r *registryConfigurator) addDNSMatchPolicy(domains []string, ip netip.Addr) error {
 	// if the gpo key is present, we need to put our DNS settings there, otherwise our config might be ignored
 	// see https://learn.microsoft.com/en-us/openspecs/windows_protocols/ms-gpnrpt/8cc31cb9-20cb-4140-9e85-3e08703b4745
 
@@ -363,19 +419,17 @@ func (r *registryConfigurator) addDNSMatchPolicy(domains []string, ip netip.Addr
 		gpoPath := fmt.Sprintf("%s-%d", gpoDnsPolicyConfigMatchPath, ruleIndex)
 
 		if err := r.configureDNSPolicy(localPath, batchDomains, ip); err != nil {
-			return ruleIndex, fmt.Errorf("configure DNS Local policy for rule %d: %w", ruleIndex, err)
+			return fmt.Errorf("configure DNS Local policy for rule %d: %w", ruleIndex, err)
 		}
 
-		// Increment immediately so the caller's cleanup path knows about this rule
-		ruleIndex++
-
 		if r.gpo {
 			if err := r.configureDNSPolicy(gpoPath, batchDomains, ip); err != nil {
-				return ruleIndex, fmt.Errorf("configure gpo DNS policy for rule %d: %w", ruleIndex-1, err)
+				return fmt.Errorf("configure gpo DNS policy for rule %d: %w", ruleIndex, err)
 			}
 		}
 
-		log.Debugf("added NRPT rule %d with %d domains", ruleIndex-1, len(batchDomains))
+		log.Debugf("added NRPT rule %d with %d domains", ruleIndex, len(batchDomains))
+		ruleIndex++
 	}
 
 	if r.gpo {
@@ -385,9 +439,45 @@ func (r *registryConfigurator) addDNSMatchPolicy(domains []string, ip netip.Addr
 	}
 
 	log.Infof("added %d NRPT rules for %d domains", ruleIndex, len(domains))
-	return ruleIndex, nil
+	return nil
 }
 
+// addDNSExemptLocalPolicy carves .local back out of the catch-all. RFC 6762
+// reserves it for multicast DNS, so forwarding those names to a unicast
+// upstream answers NXDOMAIN for hosts that do exist - printers, NAS boxes, and
+// anything else announcing itself on the link - and the answer is authoritative
+// enough that Windows stops looking. A rule naming the namespace with no
+// servers hands it back to the DNS client untouched. A more specific rule still
+// wins, so a match domain under .local keeps going through us.
+func (r *registryConfigurator) addDNSExemptLocalPolicy() error {
+	var noServers netip.Addr
+
+	if err := r.configureDNSPolicy(dnsPolicyConfigExemptLocalPath, []string{nrptLocalNamespace}, noServers); err != nil {
+		return fmt.Errorf("configure exempt policy for %s: %w", nrptLocalNamespace, err)
+	}
+
+	if r.gpo {
+		if err := r.configureDNSPolicy(gpoDnsPolicyConfigExemptLocalPath, []string{nrptLocalNamespace}, noServers); err != nil {
+			return fmt.Errorf("configure gpo exempt policy for %s: %w", nrptLocalNamespace, err)
+		}
+		if err := refreshGroupPolicy(); err != nil {
+			log.Warnf("failed to refresh group policy: %v", err)
+		}
+	}
+
+	log.Infof("added NRPT exemption for %s, leaving it to the OS resolver", nrptLocalNamespace)
+	return nil
+}
+
+// configureDNSPolicy writes one NRPT rule. An invalid ip writes an exemption
+// rule: the namespace with an empty server list, which tells the DNS client to
+// resolve those names the way it would without any rule at all.
+//
+// The empty string is the whole difference, and it has to be written: dropping
+// the value and clearing ConfigOptions instead produces a rule Windows treats
+// as a no-op, keeps out of Get-DnsClientNrptPolicy -Effective, and ignores in
+// favour of the catch-all. 0x8 says the server list is the meaningful part of
+// the rule, and an empty list then means "no server, resolve normally".
 func (r *registryConfigurator) configureDNSPolicy(policyPath string, domains []string, ip netip.Addr) error {
 	if err := removeRegistryKeyFromDNSPolicyConfig(policyPath); err != nil {
 		return fmt.Errorf("remove existing dns policy: %w", err)
@@ -407,7 +497,11 @@ func (r *registryConfigurator) configureDNSPolicy(policyPath string, domains []s
 		return fmt.Errorf("set %s: %w", dnsPolicyConfigNameKey, err)
 	}
 
-	if err := regKey.SetStringValue(dnsPolicyConfigGenericDNSServersKey, ip.String()); err != nil {
+	var servers string
+	if ip.IsValid() {
+		servers = ip.String()
+	}
+	if err := regKey.SetStringValue(dnsPolicyConfigGenericDNSServersKey, servers); err != nil {
 		return fmt.Errorf("set %s: %w", dnsPolicyConfigGenericDNSServersKey, err)
 	}
 
@@ -450,7 +544,7 @@ func (r *registryConfigurator) flushDNSCache() {
 
 	ret, _, err := dnsFlushResolverCacheFn.Call()
 	if ret == 0 {
-		if err != nil && !errors.Is(err, syscall.Errno(0)) {
+		if !errors.Is(err, syscall.Errno(0)) {
 			log.Errorf("DnsFlushResolverCache failed: %v", err)
 			return
 		}
@@ -496,7 +590,7 @@ func (r *registryConfigurator) deleteInterfaceRegistryKeyProperty(propertyKey st
 }
 
 func (r *registryConfigurator) getInterfaceRegistryKey() (registry.Key, error) {
-	regKeyPath := interfaceConfigPath + "\\" + r.guid
+	regKeyPath := InterfaceConfigPath + "\\" + r.guid
 	regKey, err := registry.OpenKey(registry.LOCAL_MACHINE, regKeyPath, registry.SET_VALUE)
 	if err != nil {
 		return regKey, fmt.Errorf("open HKEY_LOCAL_MACHINE\\%s: %w", regKeyPath, err)
@@ -505,8 +599,11 @@ func (r *registryConfigurator) getInterfaceRegistryKey() (registry.Key, error) {
 }
 
 func (r *registryConfigurator) restoreHostDNS() error {
+	// Propagated, unlike in applyDNSConfig: there we are about to write fresh
+	// rules over whatever survived, here we are leaving, and a rule left behind
+	// keeps sending every query to an address that is about to disappear.
 	if err := r.removeDNSMatchPolicies(); err != nil {
-		log.Errorf("remove dns match policies: %s", err)
+		return fmt.Errorf("remove dns match policies: %w", err)
 	}
 
 	if err := r.deleteInterfaceRegistryKeyProperty(interfaceConfigSearchListKey); err != nil {
@@ -518,28 +615,28 @@ func (r *registryConfigurator) restoreHostDNS() error {
 	return nil
 }
 
+// removeDNSMatchPolicies deletes every NRPT rule this client may have created,
+// from the local and the GPO policy store. The rules are found by enumerating
+// the registry, the only authoritative record of what was written. Cleanup must
+// not depend on a rule count: the in-memory one is scoped to a single
+// registryConfigurator and the persisted one is deleted on every clean
+// disconnect, and a rule left behind keeps resolving names over an interface
+// that is gone, until reboot discards the volatile key.
 func (r *registryConfigurator) removeDNSMatchPolicies() error {
 	var merr *multierror.Error
 
-	// Try to remove the base entries (for backward compatibility)
-	if err := removeRegistryKeyFromDNSPolicyConfig(dnsPolicyConfigMatchPath); err != nil {
-		merr = multierror.Append(merr, fmt.Errorf("remove local base entry: %w", err))
-	}
-
-	if err := removeRegistryKeyFromDNSPolicyConfig(gpoDnsPolicyConfigMatchPath); err != nil {
-		merr = multierror.Append(merr, fmt.Errorf("remove GPO base entry: %w", err))
-	}
-
-	for i := 0; i < r.nrptEntryCount; i++ {
-		localPath := fmt.Sprintf("%s-%d", dnsPolicyConfigMatchPath, i)
-		gpoPath := fmt.Sprintf("%s-%d", gpoDnsPolicyConfigMatchPath, i)
-
-		if err := removeRegistryKeyFromDNSPolicyConfig(localPath); err != nil {
-			merr = multierror.Append(merr, fmt.Errorf("remove local entry %d: %w", i, err))
+	for _, root := range []string{DNSPolicyConfigRoot, GPODNSPolicyConfigRoot} {
+		names, err := listNRPTRuleKeys(root)
+		if err != nil {
+			merr = multierror.Append(merr, fmt.Errorf("list rule keys under %s: %w", root, err))
+			continue
 		}
 
-		if err := removeRegistryKeyFromDNSPolicyConfig(gpoPath); err != nil {
-			merr = multierror.Append(merr, fmt.Errorf("remove GPO entry %d: %w", i, err))
+		for _, name := range names {
+			path := root + `\` + name
+			if err := removeRegistryKeyFromDNSPolicyConfig(path); err != nil {
+				merr = multierror.Append(merr, fmt.Errorf("remove entry %s: %w", path, err))
+			}
 		}
 	}
 
@@ -554,11 +651,52 @@ func (r *registryConfigurator) restoreUncleanShutdownDNS() error {
 	return r.restoreHostDNS()
 }
 
+// listNRPTRuleKeys returns the names of our NRPT rule keys under a policy store
+// root. An absent root holds nothing to clean up, which is the normal state of
+// the GPO store on a machine without DNS Client policy.
+func listNRPTRuleKeys(root string) ([]string, error) {
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, root, registry.ENUMERATE_SUB_KEYS)
+	switch {
+	case errors.Is(err, registry.ErrNotExist), errors.Is(err, syscall.ERROR_PATH_NOT_FOUND):
+		// the GPO store is absent on a machine without DNS client policy
+		log.Debugf("HKEY_LOCAL_MACHINE\\%s does not exist", root)
+		return nil, nil
+	case err != nil:
+		// any other failure has to reach the caller: reporting no rules would
+		// report a successful cleanup while leaving the rules in place
+		return nil, fmt.Errorf("open HKEY_LOCAL_MACHINE\\%s: %w", root, err)
+	}
+	defer closer(k)
+
+	names, err := k.ReadSubKeyNames(-1)
+	if err != nil {
+		return nil, fmt.Errorf("read subkey names: %w", err)
+	}
+
+	var ruleKeys []string
+	for _, name := range names {
+		// registry key names are case insensitive
+		if strings.HasPrefix(strings.ToLower(name), strings.ToLower(NRPTKeyPrefix)) {
+			ruleKeys = append(ruleKeys, name)
+		}
+	}
+
+	return ruleKeys, nil
+}
+
 func removeRegistryKeyFromDNSPolicyConfig(regKeyPath string) error {
 	k, err := registry.OpenKey(registry.LOCAL_MACHINE, regKeyPath, registry.QUERY_VALUE)
-	if err != nil {
-		log.Debugf("failed to open HKEY_LOCAL_MACHINE\\%s: %v", regKeyPath, err)
+	switch {
+	case errors.Is(err, registry.ErrNotExist), errors.Is(err, syscall.ERROR_PATH_NOT_FOUND):
+		// nothing to remove, which is the normal case for a rule this config
+		// never installed
+		log.Debugf("HKEY_LOCAL_MACHINE\\%s does not exist", regKeyPath)
 		return nil
+	case err != nil:
+		// anything else has to reach the caller: reporting success here would
+		// leave the rule in force while claiming it was removed, which is how a
+		// stale rule outlives the interface it points at
+		return fmt.Errorf("open HKEY_LOCAL_MACHINE\\%s: %w", regKeyPath, err)
 	}
 
 	closer(k)
@@ -585,7 +723,7 @@ func refreshGroupPolicy() error {
 	)
 
 	if ret == 0 {
-		if err != nil && !errors.Is(err, syscall.Errno(0)) {
+		if !errors.Is(err, syscall.Errno(0)) {
 			return fmt.Errorf("RefreshPolicyEx failed: %w", err)
 		}
 		return fmt.Errorf("RefreshPolicyEx failed")
@@ -594,6 +732,20 @@ func refreshGroupPolicy() error {
 	return nil
 }
 
+func parseBoolEnv(key string) bool {
+	val := os.Getenv(key)
+	if val == "" {
+		return false
+	}
+
+	parsed, err := strconv.ParseBool(val)
+	if err != nil {
+		log.Warnf("failed to parse %s=%q: %v", key, val, err)
+		return false
+	}
+	return parsed
+}
+
 func closer(closer io.Closer) {
 	if err := closer.Close(); err != nil {
 		log.Errorf("failed to close: %s", err)

client/internal/dns/host_windows_test.go

@@ -25,7 +25,7 @@ func TestNRPTEntriesCleanupOnConfigChange(t *testing.T) {
 
 	// Create a test interface registry key so updateSearchDomains doesn't fail
 	testGUID := "{12345678-1234-1234-1234-123456789ABC}"
-	interfacePath := `SYSTEM\CurrentControlSet\Services\Tcpip\Parameters\Interfaces\` + testGUID
+	interfacePath := InterfaceConfigPath + `\` + testGUID
 	testKey, _, err := registry.CreateKey(registry.LOCAL_MACHINE, interfacePath, registry.SET_VALUE)
 	require.NoError(t, err, "Should create test interface registry key")
 	testKey.Close()
@@ -56,7 +56,7 @@ func TestNRPTEntriesCleanupOnConfigChange(t *testing.T) {
 	require.NoError(t, err)
 
 	// Verify 3 NRPT rules exist
-	assert.Equal(t, 3, cfg.nrptEntryCount, "Should create 3 NRPT rules for 125 domains")
+	assert.Equal(t, 3, countNRPTRuleKeys(t), "Should create 3 NRPT rules for 125 domains")
 	for i := 0; i < 3; i++ {
 		exists, err := registryKeyExists(fmt.Sprintf("%s-%d", dnsPolicyConfigMatchPath, i))
 		require.NoError(t, err)
@@ -81,7 +81,7 @@ func TestNRPTEntriesCleanupOnConfigChange(t *testing.T) {
 	require.NoError(t, err)
 
 	// Verify first 2 NRPT rules exist
-	assert.Equal(t, 2, cfg.nrptEntryCount, "Should create 2 NRPT rules for 75 domains")
+	assert.Equal(t, 2, countNRPTRuleKeys(t), "Should create 2 NRPT rules for 75 domains")
 	for i := 0; i < 2; i++ {
 		exists, err := registryKeyExists(fmt.Sprintf("%s-%d", dnsPolicyConfigMatchPath, i))
 		require.NoError(t, err)
@@ -94,6 +94,145 @@ func TestNRPTEntriesCleanupOnConfigChange(t *testing.T) {
 	assert.False(t, exists, "NRPT rule 2 should NOT exist after reducing to 75 domains")
 }
 
+// TestNRPTCatchAllRule verifies that RouteAll adds the root namespace to the
+// match rule instead of a rule of its own, that .local is carved back out with
+// an empty server list, and that both go away when RouteAll is cleared or the
+// host DNS is restored.
+func TestNRPTCatchAllRule(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping registry integration test in short mode")
+	}
+
+	defer cleanupRegistryKeys(t)
+	cleanupRegistryKeys(t)
+
+	testIP := netip.MustParseAddr("100.64.0.1")
+	testGUID := "{12345678-1234-1234-1234-123456789ABC}"
+	interfacePath := InterfaceConfigPath + `\` + testGUID
+	testKey, _, err := registry.CreateKey(registry.LOCAL_MACHINE, interfacePath, registry.SET_VALUE)
+	require.NoError(t, err, "Should create test interface registry key")
+	require.NoError(t, testKey.Close(), "close test interface registry key")
+	defer func() {
+		assert.NoError(t, registry.DeleteKey(registry.LOCAL_MACHINE, interfacePath), "delete test interface registry key")
+	}()
+
+	cfg := &registryConfigurator{guid: testGUID}
+
+	matchOnly := HostDNSConfig{
+		ServerIP: testIP,
+		Domains:  []DomainConfig{{Domain: "example.com", MatchOnly: true}},
+	}
+	primary := HostDNSConfig{
+		ServerIP: testIP,
+		RouteAll: true,
+		Domains:  []DomainConfig{{Domain: "example.com", MatchOnly: true}},
+	}
+	firstRule := fmt.Sprintf("%s-0", dnsPolicyConfigMatchPath)
+
+	// The root namespace is not a rule of its own: it rides in the match rule,
+	// which is the point of it not being a special case.
+	require.NoError(t, cfg.applyDNSConfig(matchOnly, nil))
+	names := ruleNamespaces(t, firstRule)
+	assert.Contains(t, names, ".example.com")
+	assert.NotContains(t, names, nrptCatchAllNamespace, "a match-only config must not claim every namespace")
+
+	require.NoError(t, cfg.applyDNSConfig(primary, nil))
+	names = ruleNamespaces(t, firstRule)
+	assert.Contains(t, names, ".example.com")
+	assert.Contains(t, names, nrptCatchAllNamespace, "RouteAll should add the root namespace to the match rule")
+
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, firstRule, registry.QUERY_VALUE)
+	require.NoError(t, err)
+	servers, _, err := k.GetStringValue(dnsPolicyConfigGenericDNSServersKey)
+	require.NoError(t, err)
+	assert.Equal(t, testIP.String(), servers, "every namespace in the rule resolves through our resolver")
+	require.NoError(t, k.Close(), "close match rule key")
+
+	// .local is carved back out: RFC 6762 reserves it for mDNS, so it needs a
+	// rule of its own — it is the one rule with a different server list.
+	ek, err := registry.OpenKey(registry.LOCAL_MACHINE, dnsPolicyConfigExemptLocalPath, registry.QUERY_VALUE)
+	require.NoError(t, err, "exemption rule should exist once the root namespace is claimed")
+
+	exemptNames, _, err := ek.GetStringsValue(dnsPolicyConfigNameKey)
+	require.NoError(t, err)
+	assert.Equal(t, []string{nrptLocalNamespace}, exemptNames, "the exemption should name only the mDNS namespace")
+
+	exemptServers, _, err := ek.GetStringValue(dnsPolicyConfigGenericDNSServersKey)
+	require.NoError(t, err, "the value has to be present, empty: without it Windows drops the rule")
+	assert.Empty(t, exemptServers, "an exemption rule lists no servers")
+
+	exemptOpts, _, err := ek.GetIntegerValue(dnsPolicyConfigConfigOptionsKey)
+	require.NoError(t, err)
+	assert.EqualValues(t, dnsPolicyConfigConfigOptionsValue, exemptOpts, "same options as a normal rule; the empty server list is what makes it an exemption")
+	require.NoError(t, ek.Close(), "close exemption rule key")
+
+	require.NoError(t, cfg.applyDNSConfig(matchOnly, nil))
+	names = ruleNamespaces(t, firstRule)
+	assert.NotContains(t, names, nrptCatchAllNamespace, "clearing RouteAll should drop the root namespace")
+
+	exists, err := registryKeyExists(dnsPolicyConfigExemptLocalPath)
+	require.NoError(t, err)
+	assert.False(t, exists, "exemption rule should go with the namespace it carves out of")
+
+	require.NoError(t, cfg.applyDNSConfig(primary, nil))
+	require.NoError(t, cfg.restoreHostDNS())
+	exists, err = registryKeyExists(firstRule)
+	require.NoError(t, err)
+	assert.False(t, exists, "restore should leave no rule behind")
+}
+
+// ruleNamespaces returns the namespaces an NRPT rule key claims.
+func ruleNamespaces(t *testing.T, path string) []string {
+	t.Helper()
+	k, err := registry.OpenKey(registry.LOCAL_MACHINE, path, registry.QUERY_VALUE)
+	require.NoError(t, err, "rule key %s should exist", path)
+	defer k.Close()
+
+	names, _, err := k.GetStringsValue(dnsPolicyConfigNameKey)
+	require.NoError(t, err)
+	return names
+}
+
+// TestNRPTCatchAllRuleLegacyEnv verifies that NB_USE_LEGACY_DNS_RESOLUTION
+// leaves the root namespace unclaimed, so no rule is written for a RouteAll
+// config that carries no match domains.
+func TestNRPTCatchAllRuleLegacyEnv(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping registry integration test in short mode")
+	}
+
+	defer cleanupRegistryKeys(t)
+	cleanupRegistryKeys(t)
+
+	t.Setenv(envLegacyDNSResolution, "true")
+
+	testGUID := "{12345678-1234-1234-1234-123456789ABC}"
+	interfacePath := InterfaceConfigPath + `\` + testGUID
+	testKey, _, err := registry.CreateKey(registry.LOCAL_MACHINE, interfacePath, registry.SET_VALUE)
+	require.NoError(t, err, "Should create test interface registry key")
+	require.NoError(t, testKey.Close(), "close test interface registry key")
+	defer func() {
+		assert.NoError(t, registry.DeleteKey(registry.LOCAL_MACHINE, interfacePath), "delete test interface registry key")
+	}()
+
+	cfg := &registryConfigurator{guid: testGUID}
+	config := HostDNSConfig{
+		ServerIP: netip.MustParseAddr("100.64.0.1"),
+		RouteAll: true,
+	}
+
+	require.NoError(t, cfg.applyDNSConfig(config, nil))
+
+	// RouteAll with no match domains and the switch set leaves nothing to write.
+	exists, err := registryKeyExists(fmt.Sprintf("%s-0", dnsPolicyConfigMatchPath))
+	require.NoError(t, err)
+	assert.False(t, exists, "no rule should be written when the legacy env var is set")
+
+	exists, err = registryKeyExists(dnsPolicyConfigExemptLocalPath)
+	require.NoError(t, err)
+	assert.False(t, exists, "no exemption without a claimed root namespace")
+}
+
 func registryKeyExists(path string) (bool, error) {
 	k, err := registry.OpenKey(registry.LOCAL_MACHINE, path, registry.QUERY_VALUE)
 	if err != nil {
@@ -106,9 +245,65 @@ func registryKeyExists(path string) (bool, error) {
 	return true, nil
 }
 
+// TestNRPTCleanupWithoutRuleCount verifies that rules written by a previous run
+// are removed by a configurator that has no record of how many there are: an
+// unclean exit loses the in-memory count and a clean disconnect deletes the
+// persisted one, so cleanup cannot depend on either.
+func TestNRPTCleanupWithoutRuleCount(t *testing.T) {
+	if testing.Short() {
+		t.Skip("skipping registry integration test in short mode")
+	}
+
+	defer cleanupRegistryKeys(t)
+	cleanupRegistryKeys(t)
+
+	testIP := netip.MustParseAddr("100.64.0.1")
+
+	// 75 domains produce two indexed rules, as the current layout does
+	domains := make([]string, 75)
+	for i := range domains {
+		domains[i] = fmt.Sprintf(".domain%d.com", i+1)
+	}
+
+	previousRun := &registryConfigurator{}
+	require.NoError(t, previousRun.addDNSMatchPolicy(domains, testIP))
+
+	// the unsuffixed key an older version would have written
+	require.NoError(t, previousRun.configureDNSPolicy(dnsPolicyConfigMatchPath, []string{".legacy.example.com"}, testIP))
+
+	// a policy owned by someone else, which cleanup must not touch
+	foreignPath := DNSPolicyConfigRoot + `\DnsPolicyConfigTestForeign`
+	foreignKey, _, err := registry.CreateKey(registry.LOCAL_MACHINE, foreignPath, registry.SET_VALUE)
+	require.NoError(t, err, "Should create foreign policy key")
+	foreignKey.Close()
+	defer func() {
+		_ = registry.DeleteKey(registry.LOCAL_MACHINE, foreignPath)
+	}()
+
+	require.Equal(t, 3, countNRPTRuleKeys(t), "Should have two indexed rules and the legacy one")
+
+	// a configurator that never applied a DNS config, as one built after a
+	// restart or from a shutdown state without a count is
+	freshRun := &registryConfigurator{}
+	require.NoError(t, freshRun.removeDNSMatchPolicies())
+
+	assert.Equal(t, 0, countNRPTRuleKeys(t), "Should remove every rule left by the previous run")
+
+	exists, err := registryKeyExists(foreignPath)
+	require.NoError(t, err)
+	assert.True(t, exists, "Should not remove a policy that is not ours")
+}
+
+func countNRPTRuleKeys(t *testing.T) int {
+	t.Helper()
+
+	names, err := listNRPTRuleKeys(DNSPolicyConfigRoot)
+	require.NoError(t, err, "Should list NRPT rule keys")
+	return len(names)
+}
+
 func cleanupRegistryKeys(*testing.T) {
-	// Clean up more entries to account for batching tests with many domains
-	cfg := &registryConfigurator{nrptEntryCount: 20}
+	cfg := &registryConfigurator{}
 	_ = cfg.removeDNSMatchPolicies()
 }
 
@@ -125,7 +320,7 @@ func TestNRPTDomainBatching(t *testing.T) {
 
 	// Create a test interface registry key so updateSearchDomains doesn't fail
 	testGUID := "{12345678-1234-1234-1234-123456789ABC}"
-	interfacePath := `SYSTEM\CurrentControlSet\Services\Tcpip\Parameters\Interfaces\` + testGUID
+	interfacePath := InterfaceConfigPath + `\` + testGUID
 	testKey, _, err := registry.CreateKey(registry.LOCAL_MACHINE, interfacePath, registry.SET_VALUE)
 	require.NoError(t, err, "Should create test interface registry key")
 	testKey.Close()
@@ -193,7 +388,7 @@ func TestNRPTDomainBatching(t *testing.T) {
 			require.NoError(t, err)
 
 			// Verify that exactly expectedRuleCount rules were created
-			assert.Equal(t, tc.expectedRuleCount, cfg.nrptEntryCount,
+			assert.Equal(t, tc.expectedRuleCount, countNRPTRuleKeys(t),
 				"Should create %d NRPT rules for %d domains", tc.expectedRuleCount, tc.domainCount)
 
 			// Verify all expected rules exist

client/internal/dns/mgmt/mgmt_refresh_test.go

@@ -224,6 +224,7 @@ func TestResolver_StaleTriggersAsyncRefresh(t *testing.T) {
 }
 
 func TestResolver_ConcurrentStaleHitsCollapseRefresh(t *testing.T) {
+	semaphore := make(chan struct{})
 	r := NewResolver()
 	chain := newFakeChain()
 	chain.setAnswer("mgmt.example.com.", dns.TypeA, "10.0.0.2")
@@ -239,7 +240,7 @@ func TestResolver_ConcurrentStaleHitsCollapseRefresh(t *testing.T) {
 				break
 			}
 		}
-		time.Sleep(50 * time.Millisecond) // hold inflight long enough to collide
+		<-semaphore // block the call to force request collision
 	}
 
 	r.SetChainResolver(chain, 50)
@@ -255,17 +256,17 @@ func TestResolver_ConcurrentStaleHitsCollapseRefresh(t *testing.T) {
 
 	var wg sync.WaitGroup
 	for i := 0; i < 50; i++ {
-		wg.Add(1)
-		go func() {
-			defer wg.Done()
+		wg.Go(func() {
 			queryA(t, r, "mgmt.example.com.")
-		}()
+		})
 	}
+
+	assert.Eventually(t, func() bool { return inflight.Load() >= 1 }, 2*time.Second, 100*time.Millisecond)
+
+	close(semaphore)
 	wg.Wait()
 
-	waitFor(t, 2*time.Second, func() bool {
-		return inflight.Load() == 0
-	})
+	assert.Eventually(t, func() bool { return inflight.Load() == 0 }, 2*time.Second, 100*time.Millisecond)
 
 	calls := chain.callCount("mgmt.example.com.", dns.TypeA)
 	assert.LessOrEqual(t, calls, 2, "singleflight must collapse concurrent refreshes (got %d)", calls)

client/internal/dns/response_writer_test.go

@@ -4,7 +4,7 @@ import (
 	"net"
 	"testing"
 
-	"github.com/golang/mock/gomock"
+	"go.uber.org/mock/gomock"
 	"github.com/google/gopacket"
 	"github.com/google/gopacket/layers"
 	"github.com/miekg/dns"

client/internal/dns/server_privileged_test.go

@@ -9,7 +9,7 @@ import (
 	"os"
 	"testing"
 
-	"github.com/golang/mock/gomock"
+	"go.uber.org/mock/gomock"
 	"github.com/miekg/dns"
 	"github.com/stretchr/testify/assert"
 	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"

client/internal/dns/server_test.go

@@ -423,7 +423,7 @@ func createWgInterfaceWithBind(t *testing.T) (*iface.WGIface, error) {
 		return nil, err
 	}
 
-	pf, err := uspfilter.Create(wgIface, false, flowLogger, iface.DefaultMTU)
+	pf, err := uspfilter.Create(uspfilter.Config{IFace: wgIface, FlowLogger: flowLogger, MTU: iface.DefaultMTU})
 	if err != nil {
 		t.Fatalf("failed to create uspfilter: %v", err)
 		return nil, err

client/internal/dns/service_listener.go

@@ -6,6 +6,7 @@ import (
 	"net"
 	"net/netip"
 	"runtime"
+	"slices"
 	"strconv"
 	"sync"
 	"time"
@@ -17,17 +18,20 @@ import (
 	nberrors "github.com/netbirdio/netbird/client/errors"
 
 	firewall "github.com/netbirdio/netbird/client/firewall/manager"
-	"github.com/netbirdio/netbird/client/internal/ebpf"
-	ebpfMgr "github.com/netbirdio/netbird/client/internal/ebpf/manager"
 )
 
 const (
 	customPort = 5053
+	// randomPortAttempts bounds the search for a port free on both protocols.
+	randomPortAttempts = 5
 )
 
 var (
 	defaultIP = netip.MustParseAddr("127.0.0.1")
 	customIP  = netip.MustParseAddr("127.0.0.153")
+
+	// dnatProtocols are the protocols the port 53 redirect covers.
+	dnatProtocols = []firewall.Protocol{firewall.ProtocolUDP, firewall.ProtocolTCP}
 )
 
 type serviceViaListener struct {
@@ -40,9 +44,20 @@ type serviceViaListener struct {
 	listenPort        uint16
 	listenerIsRunning bool
 	listenerFlagLock  sync.Mutex
-	ebpfService       ebpfMgr.Manager
 	firewall          Firewall
-	tcpDNATConfigured bool
+	// dnatRules holds the port 53 redirects that are installed and not yet
+	// removed, so a removal that fails can be retried.
+	dnatRules []dnatRule
+}
+
+// dnatRule is a port 53 redirect as it was installed. The target is kept with
+// the rule because the listener can come back on a different address or port,
+// and a retried removal has to name the address and port the rule was added
+// with, not the ones in use now.
+type dnatRule struct {
+	protocol firewall.Protocol
+	ip       netip.Addr
+	port     uint16
 }
 
 func newServiceViaListener(wgIface WGIface, customAddr *netip.AddrPort, fw Firewall) *serviceViaListener {
@@ -112,34 +127,93 @@ func (s *serviceViaListener) Listen() error {
 		}
 	}()
 
-	// When eBPF redirects UDP port 53 to our listen port, TCP still needs
-	// a DNAT rule because eBPF only handles UDP.
-	if s.ebpfService != nil && s.firewall != nil && s.listenPort != DefaultPort {
-		if err := s.firewall.AddOutputDNAT(s.listenIP, firewall.ProtocolTCP, DefaultPort, s.listenPort); err != nil {
-			log.Warnf("failed to add DNS TCP DNAT rule, TCP DNS on port 53 will not work: %v", err)
-		} else {
-			s.tcpDNATConfigured = true
-			log.Infof("added DNS TCP DNAT rule: %s:%d -> %s:%d", s.listenIP, DefaultPort, s.listenIP, s.listenPort)
-		}
+	if s.listenPort != DefaultPort {
+		s.setupDNAT()
 	}
 
 	return nil
 }
 
+// setupDNAT redirects port 53 to the port the DNS server actually listens on.
+// Both protocols must be redirected or none: RuntimePort reports port 53 only
+// while the full redirect is in place, so a half-configured redirect would
+// advertise a resolver that answers over one protocol.
+func (s *serviceViaListener) setupDNAT() {
+	if s.firewall == nil {
+		log.Errorf("no firewall manager available to redirect DNS port %d to %d, "+
+			"clients pointed at %s will not reach the resolver", DefaultPort, s.listenPort, s.listenIP)
+		return
+	}
+
+	// Clear whatever an earlier removal left behind first. Those rules can point
+	// at an address or port this listener no longer uses, and they are matched
+	// before anything added now, so adding a redirect on top of one would keep
+	// sending port 53 traffic to the previous listener while reporting the
+	// redirect as complete. The rules stay recorded for a later attempt.
+	if err := s.removeDNAT(); err != nil {
+		log.Errorf("failed to remove stale DNS DNAT rules, leaving port %d redirected to the previous listener: %v",
+			DefaultPort, err)
+		return
+	}
+
+	for _, proto := range dnatProtocols {
+		if err := s.firewall.AddOutputDNAT(s.listenIP, proto, DefaultPort, s.listenPort); err != nil {
+			log.Errorf("failed to add DNS %s DNAT rule, DNS on port %d will not work: %v",
+				proto, DefaultPort, err)
+			if err := s.removeDNAT(); err != nil {
+				log.Warnf("failed to roll back DNS DNAT rules, retrying on stop: %v", err)
+			}
+			return
+		}
+		s.dnatRules = append(s.dnatRules, dnatRule{protocol: proto, ip: s.listenIP, port: s.listenPort})
+	}
+
+	log.Infof("added DNS DNAT rules: %s:%d -> %s:%d (UDP + TCP)", s.listenIP, DefaultPort, s.listenIP, s.listenPort)
+}
+
+// removeDNAT removes every installed port 53 redirect. A rule whose removal
+// fails stays recorded so a later setup or Stop retries it, rather than leaving
+// port 53 pointing at a resolver that is no longer listening.
+func (s *serviceViaListener) removeDNAT() error {
+	if s.firewall == nil {
+		return nil
+	}
+
+	var merr *multierror.Error
+	var remaining []dnatRule
+	for _, rule := range s.dnatRules {
+		if err := s.firewall.RemoveOutputDNAT(rule.ip, rule.protocol, DefaultPort, rule.port); err != nil {
+			merr = multierror.Append(merr, fmt.Errorf("remove DNS %s DNAT rule for %s:%d: %w",
+				rule.protocol, rule.ip, rule.port, err))
+			remaining = append(remaining, rule)
+		}
+	}
+	s.dnatRules = remaining
+
+	return nberrors.FormatErrorOrNil(merr)
+}
+
 func (s *serviceViaListener) Stop() error {
 	s.listenerFlagLock.Lock()
 	defer s.listenerFlagLock.Unlock()
 
+	var merr *multierror.Error
+
+	// Redirects are removed even when the listener is already stopped, so that
+	// a removal which failed earlier is retried instead of leaving port 53
+	// pointing at a resolver that no longer listens.
+	if err := s.removeDNAT(); err != nil {
+		merr = multierror.Append(merr, err)
+	}
+
 	if !s.listenerIsRunning {
-		return nil
+		return nberrors.FormatErrorOrNil(merr)
 	}
 	s.listenerIsRunning = false
 
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 
-	var merr *multierror.Error
-
 	if err := s.server.ShutdownContext(ctx); err != nil {
 		merr = multierror.Append(merr, fmt.Errorf("stop DNS UDP server: %w", err))
 	}
@@ -148,19 +222,6 @@ func (s *serviceViaListener) Stop() error {
 		merr = multierror.Append(merr, fmt.Errorf("stop DNS TCP server: %w", err))
 	}
 
-	if s.tcpDNATConfigured && s.firewall != nil {
-		if err := s.firewall.RemoveOutputDNAT(s.listenIP, firewall.ProtocolTCP, DefaultPort, s.listenPort); err != nil {
-			merr = multierror.Append(merr, fmt.Errorf("remove DNS TCP DNAT rule: %w", err))
-		}
-		s.tcpDNATConfigured = false
-	}
-
-	if s.ebpfService != nil {
-		if err := s.ebpfService.FreeDNSFwd(); err != nil {
-			merr = multierror.Append(merr, fmt.Errorf("stop traffic forwarder: %w", err))
-		}
-	}
-
 	return nberrors.FormatErrorOrNil(merr)
 }
 
@@ -177,11 +238,23 @@ func (s *serviceViaListener) RuntimePort() int {
 	s.listenerFlagLock.Lock()
 	defer s.listenerFlagLock.Unlock()
 
-	if s.ebpfService != nil {
+	if s.redirectInstalled() {
 		return DefaultPort
-	} else {
-		return int(s.listenPort)
 	}
+	return int(s.listenPort)
+}
+
+// redirectInstalled reports whether every protocol is redirected from port 53
+// to the address and port the listener currently serves. Rules left over from
+// an earlier listener do not count.
+func (s *serviceViaListener) redirectInstalled() bool {
+	for _, proto := range dnatProtocols {
+		current := dnatRule{protocol: proto, ip: s.listenIP, port: s.listenPort}
+		if !slices.Contains(s.dnatRules, current) {
+			return false
+		}
+	}
+	return true
 }
 
 func (s *serviceViaListener) RuntimeIP() netip.Addr {
@@ -190,30 +263,29 @@ func (s *serviceViaListener) RuntimeIP() netip.Addr {
 
 // evalListenAddress figures out the listen address for the DNS server.
 // IPv4-only: all peers have a v4 overlay address, and DNS config points to v4.
-// First checks port 53 on WG interface or lo, then tries eBPF on a random port,
-// then falls back to port 5053.
+// Prefers port 53 on the overlay interface or lo, so no redirect is needed at
+// all; when it is taken it falls back to port 5053 and then to a random free
+// port, both of which need the port 53 redirect set up by setupDNAT.
 func (s *serviceViaListener) evalListenAddress() (netip.Addr, uint16, error) {
 	if s.customAddr != nil {
 		return s.customAddr.Addr(), s.customAddr.Port(), nil
 	}
 
-	ip, ok := s.testFreePort(DefaultPort)
-	if ok {
+	if ip, ok := s.testFreePort(DefaultPort); ok {
 		return ip, DefaultPort, nil
 	}
 
-	ebpfSrv, port, ok := s.tryToUseeBPF()
-	if ok {
-		s.ebpfService = ebpfSrv
-		return s.wgInterface.Address().IP, port, nil
-	}
-
-	ip, ok = s.testFreePort(customPort)
-	if ok {
+	if ip, ok := s.testFreePort(customPort); ok {
 		return ip, customPort, nil
 	}
 
-	return netip.Addr{}, 0, fmt.Errorf("failed to find a free port for DNS server")
+	ip := s.wgInterface.Address().IP
+	port, err := s.randomFreePort(ip)
+	if err != nil {
+		return netip.Addr{}, 0, fmt.Errorf("find a free port for DNS server: %w", err)
+	}
+
+	return ip, port, nil
 }
 
 func (s *serviceViaListener) testFreePort(port int) (netip.Addr, bool) {
@@ -260,48 +332,25 @@ func (s *serviceViaListener) tryToBind(ip netip.Addr, port int) bool {
 	return true
 }
 
-// tryToUseeBPF decides whether to apply eBPF program to capture DNS traffic on port 53.
-// This is needed because on some operating systems if we start a DNS server not on a default port 53,
-// the domain name  resolution won't work. So, in case we are running on Linux and picked a free
-// port we should fall back to the eBPF solution that will capture traffic on port 53 and forward
-// it to a local DNS server running on the chosen port.
-func (s *serviceViaListener) tryToUseeBPF() (ebpfMgr.Manager, uint16, bool) {
-	if runtime.GOOS != "linux" {
-		return nil, 0, false
+// randomFreePort returns a port that is free on ip for both UDP and TCP, since
+// the DNS server binds both. The probe listeners are closed again, so the port
+// is only likely, not guaranteed, to still be free when the server binds it.
+func (s *serviceViaListener) randomFreePort(ip netip.Addr) (uint16, error) {
+	for range randomPortAttempts {
+		probeListener, err := net.ListenUDP("udp4", &net.UDPAddr{})
+		if err != nil {
+			return 0, fmt.Errorf("bind random port: %w", err)
+		}
+
+		port := uint16(probeListener.LocalAddr().(*net.UDPAddr).Port)
+		if err := probeListener.Close(); err != nil {
+			return 0, fmt.Errorf("free up probed port: %w", err)
+		}
+
+		if s.tryToBind(ip, int(port)) {
+			return port, nil
+		}
 	}
 
-	port, err := s.generateFreePort() //nolint:staticcheck,unused
-	if err != nil {
-		log.Warnf("failed to generate a free port for eBPF DNS forwarder server: %s", err)
-		return nil, 0, false
-	}
-
-	ebpfSrv := ebpf.GetEbpfManagerInstance()
-	err = ebpfSrv.LoadDNSFwd(s.wgInterface.Address().IP, int(port))
-	if err != nil {
-		log.Warnf("failed to load DNS forwarder eBPF program, error: %s", err)
-		return nil, 0, false
-	}
-
-	return ebpfSrv, port, true
-}
-
-func (s *serviceViaListener) generateFreePort() (uint16, error) {
-	ok := s.tryToBind(s.wgInterface.Address().IP, customPort)
-	if ok {
-		return customPort, nil
-	}
-
-	probeListener, err := net.ListenUDP("udp4", &net.UDPAddr{})
-	if err != nil {
-		log.Debugf("failed to bind random port for DNS: %s", err)
-		return 0, err
-	}
-
-	port := uint16(probeListener.LocalAddr().(*net.UDPAddr).Port)
-	if err = probeListener.Close(); err != nil {
-		log.Debugf("failed to free up DNS port: %s", err)
-		return 0, err
-	}
-	return port, nil
+	return 0, fmt.Errorf("no port free for UDP and TCP on %s after %d attempts", ip, randomPortAttempts)
 }

client/internal/dns/service_listener_test.go

@@ -1,6 +1,7 @@
 package dns
 
 import (
+	"errors"
 	"fmt"
 	"net"
 	"net/netip"
@@ -10,6 +11,8 @@ import (
 	"github.com/miekg/dns"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+
+	firewall "github.com/netbirdio/netbird/client/firewall/manager"
 )
 
 func TestServiceViaListener_TCPAndUDP(t *testing.T) {
@@ -84,3 +87,133 @@ func TestServiceViaListener_TCPAndUDP(t *testing.T) {
 	require.NotEmpty(t, tcpResp.Answer)
 	assert.Contains(t, tcpResp.Answer[0].String(), "192.0.2.1", "TCP response should contain expected IP")
 }
+
+type dnatCall struct {
+	rule  dnatRule
+	added bool
+}
+
+// fakeFirewall records DNAT calls and fails the ones named in addErrs/removeErrs.
+type fakeFirewall struct {
+	calls      []dnatCall
+	addErrs    map[firewall.Protocol]error
+	removeErrs map[firewall.Protocol]error
+}
+
+func (f *fakeFirewall) AddOutputDNAT(ip netip.Addr, protocol firewall.Protocol, _, translatedPort uint16) error {
+	if err := f.addErrs[protocol]; err != nil {
+		return err
+	}
+	f.calls = append(f.calls, dnatCall{rule: dnatRule{protocol: protocol, ip: ip, port: translatedPort}, added: true})
+	return nil
+}
+
+func (f *fakeFirewall) RemoveOutputDNAT(ip netip.Addr, protocol firewall.Protocol, _, translatedPort uint16) error {
+	if err := f.removeErrs[protocol]; err != nil {
+		return err
+	}
+	f.calls = append(f.calls, dnatCall{rule: dnatRule{protocol: protocol, ip: ip, port: translatedPort}})
+	return nil
+}
+
+func newDNATTestService(fw Firewall) *serviceViaListener {
+	return &serviceViaListener{
+		listenIP:   netip.MustParseAddr("100.64.0.1"),
+		listenPort: customPort,
+		firewall:   fw,
+	}
+}
+
+func TestSetupDNAT_BothProtocols(t *testing.T) {
+	svc := newDNATTestService(&fakeFirewall{})
+
+	svc.setupDNAT()
+
+	assert.Len(t, svc.dnatRules, len(dnatProtocols))
+	assert.Equal(t, DefaultPort, svc.RuntimePort(), "port 53 is advertised once both redirects are installed")
+}
+
+func TestSetupDNAT_RollsBackPartialRedirect(t *testing.T) {
+	fw := &fakeFirewall{addErrs: map[firewall.Protocol]error{firewall.ProtocolTCP: errors.New("nftables busy")}}
+	svc := newDNATTestService(fw)
+
+	svc.setupDNAT()
+
+	assert.Empty(t, svc.dnatRules, "the UDP redirect installed before the failure must be rolled back")
+	assert.Equal(t, int(svc.listenPort), svc.RuntimePort(), "an incomplete redirect must not advertise port 53")
+	udp := dnatRule{protocol: firewall.ProtocolUDP, ip: svc.listenIP, port: svc.listenPort}
+	assert.Contains(t, fw.calls, dnatCall{rule: udp}, "UDP removal should have been attempted")
+}
+
+// A rollback that fails must keep the rule recorded, so port 53 is not left
+// redirected to a resolver that no longer listens.
+func TestStop_RetriesFailedDNATRemoval(t *testing.T) {
+	fw := &fakeFirewall{
+		addErrs:    map[firewall.Protocol]error{firewall.ProtocolTCP: errors.New("nftables busy")},
+		removeErrs: map[firewall.Protocol]error{firewall.ProtocolUDP: errors.New("nftables busy")},
+	}
+	svc := newDNATTestService(fw)
+
+	svc.setupDNAT()
+	udp := dnatRule{protocol: firewall.ProtocolUDP, ip: svc.listenIP, port: svc.listenPort}
+	require.Equal(t, []dnatRule{udp}, svc.dnatRules, "a failed rollback keeps the rule for a later retry")
+
+	require.Error(t, svc.Stop(), "the failing removal should be reported")
+	require.Equal(t, []dnatRule{udp}, svc.dnatRules)
+
+	delete(fw.removeErrs, firewall.ProtocolUDP)
+	require.NoError(t, svc.Stop(), "a later stop retries the removal")
+	assert.Empty(t, svc.dnatRules)
+}
+
+// A stale rule that cannot be removed is matched before anything added now, so
+// no new redirect may be installed on top of it and port 53 must not be
+// advertised as reaching this listener.
+func TestSetupDNAT_AbortsWhileStaleRuleRemains(t *testing.T) {
+	fw := &fakeFirewall{removeErrs: map[firewall.Protocol]error{firewall.ProtocolUDP: errors.New("nftables busy")}}
+	svc := newDNATTestService(fw)
+	stalePort := svc.listenPort
+
+	svc.setupDNAT()
+	require.Error(t, svc.Stop())
+	staleUDP := dnatRule{protocol: firewall.ProtocolUDP, ip: svc.listenIP, port: stalePort}
+	require.Equal(t, []dnatRule{staleUDP}, svc.dnatRules)
+
+	svc.listenPort = stalePort + 1
+	fw.calls = nil
+
+	svc.setupDNAT()
+
+	assert.Equal(t, []dnatRule{staleUDP}, svc.dnatRules, "the stale rule stays recorded for a later attempt")
+	for _, call := range fw.calls {
+		assert.False(t, call.added, "no redirect may be installed while a stale one is still in place")
+	}
+	assert.Equal(t, int(svc.listenPort), svc.RuntimePort(), "port 53 must not be advertised")
+}
+
+// A rule left behind by a failed removal must be removed with the address and
+// port it was installed with, even when the listener has since moved to another
+// port, and it must not count towards the redirect the new listener advertises.
+func TestSetupDNAT_ClearsStaleRuleAfterPortChange(t *testing.T) {
+	fw := &fakeFirewall{removeErrs: map[firewall.Protocol]error{firewall.ProtocolUDP: errors.New("nftables busy")}}
+	svc := newDNATTestService(fw)
+	stalePort := svc.listenPort
+
+	svc.setupDNAT()
+	require.Error(t, svc.Stop())
+	staleUDP := dnatRule{protocol: firewall.ProtocolUDP, ip: svc.listenIP, port: stalePort}
+	require.Equal(t, []dnatRule{staleUDP}, svc.dnatRules)
+
+	delete(fw.removeErrs, firewall.ProtocolUDP)
+	svc.listenPort = stalePort + 1
+	fw.calls = nil
+
+	svc.setupDNAT()
+
+	assert.Contains(t, fw.calls, dnatCall{rule: staleUDP}, "the stale rule must be removed with its original port")
+	assert.Len(t, svc.dnatRules, len(dnatProtocols))
+	assert.Equal(t, DefaultPort, svc.RuntimePort(), "the new listener is fully redirected")
+	for _, rule := range svc.dnatRules {
+		assert.Equal(t, svc.listenPort, rule.port, "only rules for the current listener remain")
+	}
+}

client/internal/dns/unclean_shutdown_windows.go

@@ -5,9 +5,8 @@ import (
 )
 
 type ShutdownState struct {
-	Guid           string
-	GPO            bool
-	NRPTEntryCount int
+	Guid string
+	GPO  bool
 }
 
 func (s *ShutdownState) Name() string {
@@ -16,9 +15,8 @@ func (s *ShutdownState) Name() string {
 
 func (s *ShutdownState) Cleanup() error {
 	manager := &registryConfigurator{
-		guid:           s.Guid,
-		gpo:            s.GPO,
-		nrptEntryCount: s.NRPTEntryCount,
+		guid: s.Guid,
+		gpo:  s.GPO,
 	}
 
 	if err := manager.restoreUncleanShutdownDNS(); err != nil {

client/internal/dns_test.go

@@ -8,7 +8,9 @@ import (
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
 
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
 	nbdns "github.com/netbirdio/netbird/dns"
+	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
 )
 
 func TestCreatePTRRecord_IPv4(t *testing.T) {
@@ -136,3 +138,88 @@ func TestAddReverseZone_IPv6(t *testing.T) {
 	assert.Len(t, reverseZone.Records, 1)
 	assert.Equal(t, int(dns.TypePTR), reverseZone.Records[0].Type)
 }
+
+// TestToDNSConfig_ZoneFlagsPreserved pins the per-zone NonAuthoritative flag
+// through the legacy DNSConfig path. A non-authoritative zone is match-only:
+// the local resolver falls through to the upstream for an in-zone name it does
+// not define. The built-in peer zone is the authoritative one and must stay
+// that way, so the flag has to travel per zone rather than be derived.
+func TestToDNSConfig_ZoneFlagsPreserved(t *testing.T) {
+	config := toDNSConfig(&mgmProto.DNSConfig{
+		ServiceEnable: true,
+		CustomZones: []*mgmProto.CustomZone{
+			{
+				Domain: "netbird.cloud.",
+				Records: []*mgmProto.SimpleRecord{
+					{Name: "peer1.netbird.cloud.", Type: int64(dns.TypeA), Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.1"},
+				},
+			},
+			{
+				Domain:               "corp.internal.",
+				NonAuthoritative:     true,
+				SearchDomainDisabled: true,
+				Records: []*mgmProto.SimpleRecord{
+					{Name: "db.corp.internal.", Type: int64(dns.TypeA), Class: nbdns.DefaultClass, TTL: 300, RData: "10.10.0.5"},
+				},
+			},
+		},
+	}, wgaddr.Address{
+		IP:      netip.MustParseAddr("100.64.0.1"),
+		Network: netip.MustParsePrefix("100.64.0.0/16"),
+	})
+
+	zones := make(map[string]nbdns.CustomZone, len(config.CustomZones))
+	for _, zone := range config.CustomZones {
+		zones[zone.Domain] = zone
+	}
+
+	peerZone, ok := zones["netbird.cloud."]
+	require.True(t, ok, "peer zone must survive")
+	assert.False(t, peerZone.NonAuthoritative, "the built-in peer zone owns the account domain and stays authoritative")
+
+	accountZone, ok := zones["corp.internal."]
+	require.True(t, ok, "account zone must survive")
+	assert.True(t, accountZone.NonAuthoritative, "an account zone stays match-only, else undefined in-zone names get black-holed")
+	assert.True(t, accountZone.SearchDomainDisabled)
+}
+
+// TestToDNSConfig_SingleZoneForcedAuthoritative pins the compatibility clause
+// in toDNSConfig: a config carrying exactly one zone is treated as
+// authoritative no matter what the server said, because servers that predate
+// the NonAuthoritative field send only the peer FQDN zone.
+//
+// The clause can only ever downgrade an explicit true to false, so a server
+// that legitimately sends a single non-authoritative zone — an account whose
+// only zone is a custom one, with no peer records to build the built-in zone
+// from — gets that zone's whole apex black-holed on the client. Real accounts
+// always carry the peer zone alongside, which is why this is latent. Narrowing
+// it needs a way to tell "unset" from "false" on the wire, or the account
+// domain passed down here; until then this test states the contract so a
+// change to it is deliberate.
+func TestToDNSConfig_SingleZoneForcedAuthoritative(t *testing.T) {
+	config := toDNSConfig(&mgmProto.DNSConfig{
+		ServiceEnable: true,
+		CustomZones: []*mgmProto.CustomZone{
+			{
+				Domain:           "corp.internal.",
+				NonAuthoritative: true,
+				Records: []*mgmProto.SimpleRecord{
+					{Name: "db.corp.internal.", Type: int64(dns.TypeA), Class: nbdns.DefaultClass, TTL: 300, RData: "10.10.0.5"},
+				},
+			},
+		},
+	}, wgaddr.Address{
+		IP:      netip.MustParseAddr("100.64.0.1"),
+		Network: netip.MustParsePrefix("100.64.0.0/16"),
+	})
+
+	require.NotEmpty(t, config.CustomZones)
+	assert.Equal(t, "corp.internal.", config.CustomZones[0].Domain)
+	assert.False(t, config.CustomZones[0].NonAuthoritative,
+		"a lone zone is forced authoritative for pre-NonAuthoritative servers")
+
+	// The reverse zone the config gains afterwards must not feed back into the
+	// decision: the compat gate counts the zones the server sent.
+	require.Len(t, config.CustomZones, 2, "a reverse zone is appended for the overlay prefix")
+	assert.Equal(t, "64.100.in-addr.arpa.", config.CustomZones[1].Domain)
+}

client/internal/dnsfwd/forwarder.go

@@ -54,12 +54,20 @@ type DNSForwarder struct {
 	ttl            uint32
 	statusRecorder *peer.Status
 
-	dnsServer *dns.Server
-	mux       *dns.ServeMux
-	tcpServer *dns.Server
-	tcpMux    *dns.ServeMux
+	mux    *dns.ServeMux
+	tcpMux *dns.ServeMux
 
-	mutex      sync.RWMutex
+	mutex sync.RWMutex
+	// closed records that Close has run, so a Listen still in flight does not
+	// go on to serve sockets nobody will shut down.
+	closed bool
+	// The sockets are kept alongside the servers because closing them is the
+	// only stop that always works: a server whose ActivateAndServe has not run
+	// yet refuses to shut down, and would otherwise start serving afterwards.
+	udpConn    net.PacketConn
+	tcpLn      net.Listener
+	dnsServer  *dns.Server
+	tcpServer  *dns.Server
 	fwdEntries []*ForwarderEntry
 	firewall   firewaller
 	resolver   resolver
@@ -106,7 +114,7 @@ func (f *DNSForwarder) Listen(entries []*ForwarderEntry) error {
 	mux := dns.NewServeMux()
 	f.mux = mux
 	mux.HandleFunc(".", f.handleDNSQueryUDP)
-	f.dnsServer = &dns.Server{
+	dnsServer := &dns.Server{
 		PacketConn: udpLn,
 		Handler:    mux,
 	}
@@ -114,22 +122,32 @@ func (f *DNSForwarder) Listen(entries []*ForwarderEntry) error {
 	tcpMux := dns.NewServeMux()
 	f.tcpMux = tcpMux
 	tcpMux.HandleFunc(".", f.handleDNSQueryTCP)
-	f.tcpServer = &dns.Server{
+	tcpServer := &dns.Server{
 		Listener: tcpLn,
 		Handler:  tcpMux,
 	}
 
-	f.UpdateDomains(entries)
+	if !f.publish(udpLn, tcpLn, dnsServer, tcpServer, entries) {
+		log.Infof("DNS forwarder on %s was closed before it started serving", addrDesc)
+		if err := udpLn.Close(); err != nil {
+			log.Debugf("close UDP listener of a closed forwarder: %v", err)
+		}
+		if err := tcpLn.Close(); err != nil {
+			log.Debugf("close TCP listener of a closed forwarder: %v", err)
+		}
+		return nil
+	}
+	log.Debugf("DNS forwarder serving %d domains", len(entries))
 
 	errCh := make(chan error, 2)
 
 	go func() {
 		log.Infof("DNS UDP listener running on %s", addrDesc)
-		errCh <- f.dnsServer.ActivateAndServe()
+		errCh <- dnsServer.ActivateAndServe()
 	}()
 	go func() {
 		log.Infof("DNS TCP listener running on %s", addrDesc)
-		errCh <- f.tcpServer.ActivateAndServe()
+		errCh <- tcpServer.ActivateAndServe()
 	}()
 
 	return <-errCh
@@ -151,6 +169,46 @@ func (f *DNSForwarder) createTCPListener(netstackNet *netstack.Net) (net.Listene
 	return net.ListenTCP("tcp", net.TCPAddrFromAddrPort(f.listenAddress))
 }
 
+// publish hands the sockets, servers and entries to the forwarder so Close can
+// reach them and Domains can report them, and says whether serving may begin.
+// Listen runs on its own goroutine, so a Close can arrive before it gets this
+// far; false means the caller must close what it created instead of serving on
+// it.
+//
+// The entries go in under the same lock rather than afterwards. Anything that
+// reads them in between would otherwise see a forwarder that is listening and
+// serves no domain, which for a caller rebuilding one means it comes back
+// refusing every routed query.
+func (f *DNSForwarder) publish(
+	udpConn net.PacketConn,
+	tcpLn net.Listener,
+	dnsServer, tcpServer *dns.Server,
+	entries []*ForwarderEntry,
+) bool {
+	f.mutex.Lock()
+	defer f.mutex.Unlock()
+
+	if f.closed {
+		return false
+	}
+
+	f.udpConn = udpConn
+	f.tcpLn = tcpLn
+	f.dnsServer = dnsServer
+	f.tcpServer = tcpServer
+	f.fwdEntries = entries
+	return true
+}
+
+// Domains returns the entries currently being served. The slice is replaced
+// wholesale by UpdateDomains rather than mutated, so the caller may read it but
+// must not write to it.
+func (f *DNSForwarder) Domains() []*ForwarderEntry {
+	f.mutex.RLock()
+	defer f.mutex.RUnlock()
+	return f.fwdEntries
+}
+
 func (f *DNSForwarder) UpdateDomains(entries []*ForwarderEntry) {
 	f.mutex.Lock()
 	defer f.mutex.Unlock()
@@ -189,19 +247,45 @@ func (f *DNSForwarder) removeStaleCacheEntries(oldEntries, newEntries []*Forward
 }
 
 func (f *DNSForwarder) Close(ctx context.Context) error {
+	// Marked closed under the lock so a Listen that has not published its
+	// servers yet gives up instead of racing this shutdown. The shutdowns
+	// themselves block, so they run outside it.
+	f.mutex.Lock()
+	f.closed = true
+	dnsServer, tcpServer := f.dnsServer, f.tcpServer
+	udpConn, tcpLn := f.udpConn, f.tcpLn
+	f.mutex.Unlock()
+
 	var result *multierror.Error
 
-	if f.dnsServer != nil {
-		if err := f.dnsServer.ShutdownContext(ctx); err != nil {
+	if dnsServer != nil {
+		if err := shutdownServer(ctx, dnsServer); err != nil {
 			result = multierror.Append(result, fmt.Errorf("UDP shutdown: %w", err))
 		}
 	}
-	if f.tcpServer != nil {
-		if err := f.tcpServer.ShutdownContext(ctx); err != nil {
+	if tcpServer != nil {
+		if err := shutdownServer(ctx, tcpServer); err != nil {
 			result = multierror.Append(result, fmt.Errorf("TCP shutdown: %w", err))
 		}
 	}
 
+	// The sockets are closed even when the shutdowns above reported nothing to
+	// do. A server that has been published but has not reached
+	// ActivateAndServe refuses to shut down, and closing what it was about to
+	// serve on is what stops it: the alternative is a listener still answering
+	// on an interface that has gone away. A shutdown that did run has already
+	// closed these, so the second close is expected to fail.
+	if udpConn != nil {
+		if err := udpConn.Close(); err != nil {
+			log.Debugf("close UDP socket of the DNS forwarder: %v", err)
+		}
+	}
+	if tcpLn != nil {
+		if err := tcpLn.Close(); err != nil {
+			log.Debugf("close TCP socket of the DNS forwarder: %v", err)
+		}
+	}
+
 	return nberrors.FormatErrorOrNil(result)
 }
 
@@ -514,3 +598,16 @@ func attachEDE(resp *dns.Msg, code uint16, text string) {
 	}
 	opt.Option = append(opt.Option, &dns.EDNS0_EDE{InfoCode: code, ExtraText: text})
 }
+
+// shutdownServer shuts a server down gracefully, treating "never started" as
+// success. A server that was published but has not reached ActivateAndServe
+// has nothing to wind down, and the caller closes its socket regardless, which
+// is what actually stops it. dns exports no sentinel for this, so the message
+// is all there is to match on.
+func shutdownServer(ctx context.Context, server *dns.Server) error {
+	err := server.ShutdownContext(ctx)
+	if err == nil || strings.Contains(err.Error(), "server not started") {
+		return nil
+	}
+	return err
+}

client/internal/dnsfwd/forwarder_test.go

@@ -1238,3 +1238,55 @@ func TestDNSForwarder_EmptyQuery(t *testing.T) {
 
 	assert.Nil(t, mockWriter.GetLastResponse(), "Should not write response for empty query")
 }
+
+// TestDNSForwarder_ClosedBeforeItServes covers Listen reaching the point of
+// serving after the forwarder has already been closed. Listen runs on its own
+// goroutine, so it can get there late, and a socket it starts serving then is
+// one nothing will ever close: on Android it keeps answering on an interface
+// that has been replaced. The close is sequenced first here rather than raced,
+// which pins the same state deterministically.
+func TestDNSForwarder_ClosedBeforeItServes(t *testing.T) {
+	f := NewDNSForwarder(netip.MustParseAddrPort("127.0.0.1:0"), 60, nil, nil, nil)
+
+	require.NoError(t, f.Close(context.Background()), "closing a forwarder that never started")
+
+	done := make(chan error, 1)
+	go func() { done <- f.Listen(nil) }()
+
+	select {
+	case err := <-done:
+		assert.NoError(t, err, "a closed forwarder should give up quietly, not serve")
+	case <-time.After(5 * time.Second):
+		t.Fatal("Listen went on to serve after the forwarder was closed")
+	}
+}
+
+// TestDNSForwarder_CloseStopsUnactivatedServers covers the window between
+// Listen publishing its servers and reaching ActivateAndServe. A server that
+// has not been activated refuses to shut down, so Close has to close the
+// sockets itself or they are left serving.
+func TestDNSForwarder_CloseStopsUnactivatedServers(t *testing.T) {
+	f := NewDNSForwarder(netip.MustParseAddrPort("127.0.0.1:0"), 60, nil, nil, nil)
+
+	udpConn, err := f.createUDPListener(nil)
+	require.NoError(t, err, "create UDP listener")
+	tcpLn, err := f.createTCPListener(nil)
+	require.NoError(t, err, "create TCP listener")
+
+	// Published but deliberately never activated, which is the state Listen is
+	// in for the moment before it starts serving.
+	require.True(t, f.publish(udpConn, tcpLn, &dns.Server{PacketConn: udpConn}, &dns.Server{Listener: tcpLn}, nil),
+		"publishing to an open forwarder")
+
+	tcpAddr := tcpLn.Addr().String()
+	require.NoError(t, f.Close(context.Background()), "close should report no error for servers it could not shut down")
+
+	_, err = tcpLn.Accept()
+	assert.Error(t, err, "the TCP socket should be closed after Close")
+
+	conn, err := net.DialTimeout("tcp", tcpAddr, time.Second)
+	if err == nil {
+		_ = conn.Close()
+		t.Fatal("the forwarder is still accepting connections after Close")
+	}
+}

client/internal/dnsfwd/manager.go

@@ -3,7 +3,6 @@ package dnsfwd
 import (
 	"context"
 	"fmt"
-	"net"
 	"net/netip"
 	"os"
 	"strconv"
@@ -101,7 +100,7 @@ func (m *Manager) Start(fwdEntries []*ForwarderEntry) error {
 	m.dnsForwarder = NewDNSForwarder(listenAddress, dnsTTL, m.firewall, m.statusRecorder, m.wgIface)
 
 	go func() {
-		if err := m.dnsForwarder.Listen(fwdEntries); err != nil {
+		if err := m.dnsForwarder.Listen(fwdEntries); err != nil { //nolint:staticcheck
 			// todo handle close error if it is exists
 			log.Errorf("failed to start DNS forwarder, err: %v", err)
 		}
@@ -118,6 +117,16 @@ func (m *Manager) UpdateDomains(entries []*ForwarderEntry) {
 	m.dnsForwarder.UpdateDomains(entries)
 }
 
+// Domains returns the entries currently being served, or nil when the
+// forwarder is not running.
+func (m *Manager) Domains() []*ForwarderEntry {
+	if m.dnsForwarder == nil {
+		return nil
+	}
+
+	return m.dnsForwarder.Domains()
+}
+
 func (m *Manager) Stop(ctx context.Context) error {
 	if m.dnsForwarder == nil {
 		return nil
@@ -160,12 +169,13 @@ func (m *Manager) allowDNSFirewall() error {
 		return nil
 	}
 
-	dnsRules, err := m.firewall.AddPeerFiltering(nil, net.IP{0, 0, 0, 0}, firewall.ProtocolUDP, nil, dport, firewall.ActionAccept, "")
+	anyV4 := []netip.Prefix{netip.PrefixFrom(netip.IPv4Unspecified(), 0)}
+	dnsRule, err := m.firewall.AddFilterRule(nil, anyV4, firewall.Network{}, firewall.ProtocolUDP, nil, dport, firewall.ActionAccept)
 	if err != nil {
 		return fmt.Errorf("add udp firewall rule: %w", err)
 	}
 
-	tcpRules, err := m.firewall.AddPeerFiltering(nil, net.IP{0, 0, 0, 0}, firewall.ProtocolTCP, nil, dport, firewall.ActionAccept, "")
+	tcpRule, err := m.firewall.AddFilterRule(nil, anyV4, firewall.Network{}, firewall.ProtocolTCP, nil, dport, firewall.ActionAccept)
 	if err != nil {
 		return fmt.Errorf("add tcp firewall rule: %w", err)
 	}
@@ -174,8 +184,12 @@ func (m *Manager) allowDNSFirewall() error {
 		return fmt.Errorf("flush: %w", err)
 	}
 
-	m.fwRules = dnsRules
-	m.tcpRules = tcpRules
+	if dnsRule != nil {
+		m.fwRules = []firewall.Rule{dnsRule}
+	}
+	if tcpRule != nil {
+		m.tcpRules = []firewall.Rule{tcpRule}
+	}
 
 	m.registerNetstackServices()
 
@@ -209,12 +223,12 @@ func (m *Manager) unregisterNetstackServices() {
 func (m *Manager) dropDNSFirewall() error {
 	var mErr *multierror.Error
 	for _, rule := range m.fwRules {
-		if err := m.firewall.DeletePeerRule(rule); err != nil {
+		if err := m.firewall.DeleteFilterRule(rule); err != nil {
 			mErr = multierror.Append(mErr, fmt.Errorf("failed to delete DNS router rules, err: %v", err))
 		}
 	}
 	for _, rule := range m.tcpRules {
-		if err := m.firewall.DeletePeerRule(rule); err != nil {
+		if err := m.firewall.DeleteFilterRule(rule); err != nil {
 			mErr = multierror.Append(mErr, fmt.Errorf("failed to delete DNS router rules, err: %v", err))
 		}
 	}

client/internal/ebpf/ebpf/bpf_bpfeb.go

@@ -1,5 +1,5 @@
 // Code generated by bpf2go; DO NOT EDIT.
-//go:build arm64be || armbe || mips || mips64 || mips64p32 || ppc64 || s390 || s390x || sparc || sparc64
+//go:build mips || mips64 || ppc64 || s390x
 
 package ebpf
 
@@ -47,9 +47,10 @@ func loadBpfObjects(obj interface{}, opts *ebpf.CollectionOptions) error {
 type bpfSpecs struct {
 	bpfProgramSpecs
 	bpfMapSpecs
+	bpfVariableSpecs
 }
 
-// bpfSpecs contains programs before they are loaded into the kernel.
+// bpfProgramSpecs contains programs before they are loaded into the kernel.
 //
 // It can be passed ebpf.CollectionSpec.Assign.
 type bpfProgramSpecs struct {
@@ -61,17 +62,28 @@ type bpfProgramSpecs struct {
 // It can be passed ebpf.CollectionSpec.Assign.
 type bpfMapSpecs struct {
 	NbFeatures           *ebpf.MapSpec `ebpf:"nb_features"`
-	NbMapDnsIp           *ebpf.MapSpec `ebpf:"nb_map_dns_ip"`
-	NbMapDnsPort         *ebpf.MapSpec `ebpf:"nb_map_dns_port"`
 	NbWgProxySettingsMap *ebpf.MapSpec `ebpf:"nb_wg_proxy_settings_map"`
 }
 
+// bpfVariableSpecs contains global variables before they are loaded into the kernel.
+//
+// It can be passed ebpf.CollectionSpec.Assign.
+type bpfVariableSpecs struct {
+	FlagFeatureWgProxy *ebpf.VariableSpec `ebpf:"flag_feature_wg_proxy"`
+	MapKeyFeatures     *ebpf.VariableSpec `ebpf:"map_key_features"`
+	MapKeyProxyPort    *ebpf.VariableSpec `ebpf:"map_key_proxy_port"`
+	MapKeyWgPort       *ebpf.VariableSpec `ebpf:"map_key_wg_port"`
+	ProxyPort          *ebpf.VariableSpec `ebpf:"proxy_port"`
+	WgPort             *ebpf.VariableSpec `ebpf:"wg_port"`
+}
+
 // bpfObjects contains all objects after they have been loaded into the kernel.
 //
 // It can be passed to loadBpfObjects or ebpf.CollectionSpec.LoadAndAssign.
 type bpfObjects struct {
 	bpfPrograms
 	bpfMaps
+	bpfVariables
 }
 
 func (o *bpfObjects) Close() error {
@@ -86,20 +98,28 @@ func (o *bpfObjects) Close() error {
 // It can be passed to loadBpfObjects or ebpf.CollectionSpec.LoadAndAssign.
 type bpfMaps struct {
 	NbFeatures           *ebpf.Map `ebpf:"nb_features"`
-	NbMapDnsIp           *ebpf.Map `ebpf:"nb_map_dns_ip"`
-	NbMapDnsPort         *ebpf.Map `ebpf:"nb_map_dns_port"`
 	NbWgProxySettingsMap *ebpf.Map `ebpf:"nb_wg_proxy_settings_map"`
 }
 
 func (m *bpfMaps) Close() error {
 	return _BpfClose(
 		m.NbFeatures,
-		m.NbMapDnsIp,
-		m.NbMapDnsPort,
 		m.NbWgProxySettingsMap,
 	)
 }
 
+// bpfVariables contains all global variables after they have been loaded into the kernel.
+//
+// It can be passed to loadBpfObjects or ebpf.CollectionSpec.LoadAndAssign.
+type bpfVariables struct {
+	FlagFeatureWgProxy *ebpf.Variable `ebpf:"flag_feature_wg_proxy"`
+	MapKeyFeatures     *ebpf.Variable `ebpf:"map_key_features"`
+	MapKeyProxyPort    *ebpf.Variable `ebpf:"map_key_proxy_port"`
+	MapKeyWgPort       *ebpf.Variable `ebpf:"map_key_wg_port"`
+	ProxyPort          *ebpf.Variable `ebpf:"proxy_port"`
+	WgPort             *ebpf.Variable `ebpf:"wg_port"`
+}
+
 // bpfPrograms contains all programs after they have been loaded into the kernel.
 //
 // It can be passed to loadBpfObjects or ebpf.CollectionSpec.LoadAndAssign.

client/internal/ebpf/ebpf/bpf_bpfel.go

@@ -1,5 +1,5 @@
 // Code generated by bpf2go; DO NOT EDIT.
-//go:build 386 || amd64 || amd64p32 || arm || arm64 || loong64 || mips64le || mips64p32le || mipsle || ppc64le || riscv64
+//go:build 386 || amd64 || arm || arm64 || loong64 || mips64le || mipsle || ppc64le || riscv64 || wasm
 
 package ebpf
 
@@ -47,9 +47,10 @@ func loadBpfObjects(obj interface{}, opts *ebpf.CollectionOptions) error {
 type bpfSpecs struct {
 	bpfProgramSpecs
 	bpfMapSpecs
+	bpfVariableSpecs
 }
 
-// bpfSpecs contains programs before they are loaded into the kernel.
+// bpfProgramSpecs contains programs before they are loaded into the kernel.
 //
 // It can be passed ebpf.CollectionSpec.Assign.
 type bpfProgramSpecs struct {
@@ -61,17 +62,28 @@ type bpfProgramSpecs struct {
 // It can be passed ebpf.CollectionSpec.Assign.
 type bpfMapSpecs struct {
 	NbFeatures           *ebpf.MapSpec `ebpf:"nb_features"`
-	NbMapDnsIp           *ebpf.MapSpec `ebpf:"nb_map_dns_ip"`
-	NbMapDnsPort         *ebpf.MapSpec `ebpf:"nb_map_dns_port"`
 	NbWgProxySettingsMap *ebpf.MapSpec `ebpf:"nb_wg_proxy_settings_map"`
 }
 
+// bpfVariableSpecs contains global variables before they are loaded into the kernel.
+//
+// It can be passed ebpf.CollectionSpec.Assign.
+type bpfVariableSpecs struct {
+	FlagFeatureWgProxy *ebpf.VariableSpec `ebpf:"flag_feature_wg_proxy"`
+	MapKeyFeatures     *ebpf.VariableSpec `ebpf:"map_key_features"`
+	MapKeyProxyPort    *ebpf.VariableSpec `ebpf:"map_key_proxy_port"`
+	MapKeyWgPort       *ebpf.VariableSpec `ebpf:"map_key_wg_port"`
+	ProxyPort          *ebpf.VariableSpec `ebpf:"proxy_port"`
+	WgPort             *ebpf.VariableSpec `ebpf:"wg_port"`
+}
+
 // bpfObjects contains all objects after they have been loaded into the kernel.
 //
 // It can be passed to loadBpfObjects or ebpf.CollectionSpec.LoadAndAssign.
 type bpfObjects struct {
 	bpfPrograms
 	bpfMaps
+	bpfVariables
 }
 
 func (o *bpfObjects) Close() error {
@@ -86,20 +98,28 @@ func (o *bpfObjects) Close() error {
 // It can be passed to loadBpfObjects or ebpf.CollectionSpec.LoadAndAssign.
 type bpfMaps struct {
 	NbFeatures           *ebpf.Map `ebpf:"nb_features"`
-	NbMapDnsIp           *ebpf.Map `ebpf:"nb_map_dns_ip"`
-	NbMapDnsPort         *ebpf.Map `ebpf:"nb_map_dns_port"`
 	NbWgProxySettingsMap *ebpf.Map `ebpf:"nb_wg_proxy_settings_map"`
 }
 
 func (m *bpfMaps) Close() error {
 	return _BpfClose(
 		m.NbFeatures,
-		m.NbMapDnsIp,
-		m.NbMapDnsPort,
 		m.NbWgProxySettingsMap,
 	)
 }
 
+// bpfVariables contains all global variables after they have been loaded into the kernel.
+//
+// It can be passed to loadBpfObjects or ebpf.CollectionSpec.LoadAndAssign.
+type bpfVariables struct {
+	FlagFeatureWgProxy *ebpf.Variable `ebpf:"flag_feature_wg_proxy"`
+	MapKeyFeatures     *ebpf.Variable `ebpf:"map_key_features"`
+	MapKeyProxyPort    *ebpf.Variable `ebpf:"map_key_proxy_port"`
+	MapKeyWgPort       *ebpf.Variable `ebpf:"map_key_wg_port"`
+	ProxyPort          *ebpf.Variable `ebpf:"proxy_port"`
+	WgPort             *ebpf.Variable `ebpf:"wg_port"`
+}
+
 // bpfPrograms contains all programs after they have been loaded into the kernel.
 //
 // It can be passed to loadBpfObjects or ebpf.CollectionSpec.LoadAndAssign.

client/internal/ebpf/ebpf/dns_fwd_linux.go

@@ -1,52 +0,0 @@
-package ebpf
-
-import (
-	"encoding/binary"
-	"fmt"
-	"net/netip"
-
-	log "github.com/sirupsen/logrus"
-)
-
-const (
-	mapKeyDNSIP   uint32 = 0
-	mapKeyDNSPort uint32 = 1
-)
-
-func (tf *GeneralManager) LoadDNSFwd(ip netip.Addr, dnsPort int) error {
-	log.Debugf("load eBPF DNS forwarder, watching addr: %s:53, redirect to port: %d", ip, dnsPort)
-	tf.lock.Lock()
-	defer tf.lock.Unlock()
-
-	err := tf.loadXdp()
-	if err != nil {
-		return err
-	}
-
-	if !ip.Is4() {
-		return fmt.Errorf("eBPF DNS forwarder only supports IPv4, got %s", ip)
-	}
-	ip4 := ip.As4()
-	err = tf.bpfObjs.NbMapDnsIp.Put(mapKeyDNSIP, binary.BigEndian.Uint32(ip4[:]))
-	if err != nil {
-		return err
-	}
-
-	err = tf.bpfObjs.NbMapDnsPort.Put(mapKeyDNSPort, uint16(dnsPort))
-	if err != nil {
-		return err
-	}
-
-	tf.setFeatureFlag(featureFlagDnsForwarder)
-	err = tf.bpfObjs.NbFeatures.Put(mapKeyFeatures, tf.featureFlags)
-	if err != nil {
-		return err
-	}
-	return nil
-}
-
-func (tf *GeneralManager) FreeDNSFwd() error {
-	log.Debugf("free ebpf DNS forwarder")
-	return tf.unsetFeatureFlag(featureFlagDnsForwarder)
-}
-

client/internal/ebpf/ebpf/manager_linux.go

@@ -15,8 +15,7 @@ import (
 const (
 	mapKeyFeatures uint32 = 0
 
-	featureFlagWGProxy      = 0b00000001
-	featureFlagDnsForwarder = 0b00000010
+	featureFlagWGProxy = 0b00000001
 )
 
 var (
@@ -28,9 +27,9 @@ var (
 
 // GeneralManager is used to load multiple eBPF programs with a custom check (if then) done in prog.c
 // The manager simply adds a feature (byte) of each program to a map that is shared between the userspace and kernel.
-// When packet arrives, the C code checks for each feature (if it is set) and executes each enabled program (e.g., dns_fwd.c and wg_proxy.c).
+// When packet arrives, the C code checks for each feature (if it is set) and executes each enabled program (e.g., wg_proxy.c).
 //
-//go:generate go run github.com/cilium/ebpf/cmd/bpf2go -cc clang-14 bpf src/prog.c -- -I /usr/x86_64-linux-gnu/include
+//go:generate go run github.com/cilium/ebpf/cmd/bpf2go -cc clang-14 bpf src/prog.c -- -I /usr/x86_64-linux-gnu/include -include src/bpf_map_def.h
 type GeneralManager struct {
 	lock         sync.Mutex
 	link         link.Link

client/internal/ebpf/ebpf/manager_linux_test.go

@@ -7,33 +7,24 @@ import (
 func TestManager_setFeatureFlag(t *testing.T) {
 	mgr := GeneralManager{}
 	mgr.setFeatureFlag(featureFlagWGProxy)
-	if mgr.featureFlags != 1 {
+	if mgr.featureFlags != featureFlagWGProxy {
 		t.Errorf("invalid feature state")
 	}
 
-	mgr.setFeatureFlag(featureFlagDnsForwarder)
-	if mgr.featureFlags != 3 {
-		t.Errorf("invalid feature state")
+	mgr.setFeatureFlag(featureFlagWGProxy)
+	if mgr.featureFlags != featureFlagWGProxy {
+		t.Errorf("setting a flag twice must be idempotent, got: %d", mgr.featureFlags)
 	}
 }
 
 func TestManager_unsetFeatureFlag(t *testing.T) {
 	mgr := GeneralManager{}
 	mgr.setFeatureFlag(featureFlagWGProxy)
-	mgr.setFeatureFlag(featureFlagDnsForwarder)
 
 	err := mgr.unsetFeatureFlag(featureFlagWGProxy)
 	if err != nil {
 		t.Errorf("unexpected error: %s", err)
 	}
-	if mgr.featureFlags != 2 {
-		t.Errorf("invalid feature state, expected: %d, got: %d", 2, mgr.featureFlags)
-	}
-
-	err = mgr.unsetFeatureFlag(featureFlagDnsForwarder)
-	if err != nil {
-		t.Errorf("unexpected error: %s", err)
-	}
 	if mgr.featureFlags != 0 {
 		t.Errorf("invalid feature state, expected: %d, got: %d", 0, mgr.featureFlags)
 	}

client/internal/ebpf/ebpf/src/bpf_map_def.h

@@ -0,0 +1,16 @@
+// libbpf 1.0 removed struct bpf_map_def, but the programs here keep the legacy
+// map definitions: they load on kernels built without BTF, which BTF-style
+// (SEC(".maps")) definitions do not. Define the struct ourselves so the
+// programs compile against current libbpf headers.
+#ifndef NB_BPF_MAP_DEF_H
+#define NB_BPF_MAP_DEF_H
+
+struct bpf_map_def {
+	unsigned int type;
+	unsigned int key_size;
+	unsigned int value_size;
+	unsigned int max_entries;
+	unsigned int map_flags;
+};
+
+#endif

client/internal/ebpf/ebpf/src/dns_fwd.c

@@ -1,67 +0,0 @@
-const __u32 map_key_dns_ip = 0;
-const __u32 map_key_dns_port = 1;
-
-struct bpf_map_def SEC("maps") nb_map_dns_ip = {
-	.type = BPF_MAP_TYPE_ARRAY,
-	.key_size = sizeof(__u32),
-	.value_size = sizeof(__u32),
-	.max_entries = 10,
-};
-
-struct bpf_map_def SEC("maps") nb_map_dns_port = {
-	.type = BPF_MAP_TYPE_ARRAY,
-	.key_size = sizeof(__u32),
-	.value_size = sizeof(__u16),
-	.max_entries = 10,
-};
-
-__be32 dns_ip = 0;
-__be16 dns_port = 0;
-
-// 13568 is 53 in big endian
-__be16 GENERAL_DNS_PORT = 13568;
-
-bool read_settings() {
-    __u16 *port_value;
-    __u32 *ip_value;
-
-    // read dns ip
-    ip_value = bpf_map_lookup_elem(&nb_map_dns_ip, &map_key_dns_ip);
-    if(!ip_value) {
-        return false;
-    }
-    dns_ip = htonl(*ip_value);
-
-    // read dns port
-    port_value = bpf_map_lookup_elem(&nb_map_dns_port, &map_key_dns_port);
-    if (!port_value) {
-        return false;
-    }
-    dns_port = htons(*port_value);
-    return true;
-}
-
-int xdp_dns_fwd(struct iphdr  *ip, struct udphdr *udp) {
-    if (dns_port == 0) {
-        if(!read_settings()){
-            return XDP_PASS;
-        }
-        // bpf_printk("dns port: %d", ntohs(dns_port));
-        // bpf_printk("dns ip: %d", ntohl(dns_ip));
-    }
-
-    if (udp->dest == GENERAL_DNS_PORT && ip->daddr == dns_ip) {
-        udp->dest = dns_port;
-        // Clear the now-stale checksum; zero means "not computed" for IPv4.
-        udp->check = 0;
-        return XDP_PASS;
-    }
-
-    if (udp->source == dns_port && ip->saddr == dns_ip) {
-        udp->source = GENERAL_DNS_PORT;
-        udp->check = 0;
-        return XDP_PASS;
-    }
-
-    return XDP_PASS;
-}

client/internal/ebpf/ebpf/src/prog.c

@@ -5,11 +5,9 @@
 #include <netinet/in.h>
 #include <linux/bpf.h>
 #include <bpf/bpf_helpers.h>
-#include "dns_fwd.c"
 #include "wg_proxy.c"
 
 const __u16 flag_feature_wg_proxy = 0b01;
-const __u16 flag_feature_dns_fwd = 0b10;
 
 const __u32 map_key_features = 0;
 struct bpf_map_def SEC("maps") nb_features = {
@@ -48,10 +46,6 @@ int nb_xdp_prog(struct xdp_md *ctx) {
        return XDP_PASS;
     }
 
-    if (*features & flag_feature_dns_fwd) {
-        xdp_dns_fwd(ip, udp);
-    }
-
     if (*features & flag_feature_wg_proxy) {
         xdp_wg_proxy(ip, udp);
     }

client/internal/ebpf/ebpf/src/readme.md

@@ -1,8 +1,18 @@
-# DNS forwarder
+# XDP programs
 
-The agent attach the XDP program to the lo device. We can not use fake address in eBPF because the
-traffic does not appear in the eBPF program. The program capture the traffic on wg_ip:53 and 
-overwrite in it the destination port to 5053.
+`prog.c` is attached to the `lo` device and dispatches to the features enabled in the
+`nb_features` map. The only feature is the WireGuard proxy (`wg_proxy.c`): it rewrites
+loopback UDP sent from the WireGuard listen port so it reaches the userspace relay proxy
+port instead, and swaps the peer endpoint port into the source so the proxy can tell
+peers apart.
+
+Maps use the legacy `struct bpf_map_def` form, defined in `bpf_map_def.h` because libbpf
+1.0 removed it. They load on kernels built without BTF, which BTF-style (`SEC(".maps")`)
+definitions do not.
+
+Regenerate the objects with `go generate ./client/internal/ebpf/ebpf/`; it needs
+`clang-14`. Loading a regenerated object needs root, attaching it needs `bpf_link`
+(kernel >= 5.7), and only one XDP program can own `lo` at a time.
 
 # Debug
 

client/internal/ebpf/manager/manager.go

@@ -1,11 +1,7 @@
 package manager
 
-import "net/netip"
-
-// Manager is used to load multiple eBPF programs. E.g., current DNS programs and WireGuard proxy
+// Manager is used to load multiple eBPF programs. E.g., the WireGuard proxy
 type Manager interface {
-	LoadDNSFwd(ip netip.Addr, dnsPort int) error
-	FreeDNSFwd() error
 	LoadWgProxy(proxyPort, wgPort int) error
 	FreeWGProxy() error
 }

client/internal/elevate/elevate.go

@@ -0,0 +1,74 @@
+// Package elevate re-runs this very executable under the operating system's own
+// privilege-elevation mechanism and waits for it to finish.
+//
+// It exists so that a change the daemon restricts to root/administrator can be
+// authorized from the GUI, by the user, at the moment they ask for it: Windows
+// shows the UAC consent dialog, macOS the system authentication dialog, and
+// Linux/FreeBSD the session's polkit agent. The credentials, where any are
+// asked for, are collected by the operating system and never pass through
+// NetBird.
+//
+// What the elevated process then does is the caller's business: it is the same
+// binary, in a one-shot mode, and it is authorized by the daemon exactly like
+// any other privileged caller, from the identity the kernel reports on the
+// control channel. Nothing here grants privilege, and the daemon gains no new
+// way to be talked into something: elevation only changes who is calling it.
+package elevate
+
+import (
+	"context"
+	"errors"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// AppliedMarker is what the elevated process prints on standard output once it has
+// done what it was run for.
+//
+// macOS's AuthorizationExecuteWithPrivileges reports no exit status and does not
+// say which process it started, so there this line is the only evidence that the
+// change was applied. The other platforms have an exit code and ignore it.
+const AppliedMarker = "netbird-elevated: applied"
+
+var (
+	// ErrDeclined reports that the user dismissed the prompt or did not
+	// authenticate. Nothing happened and nothing is wrong: a caller undoes its
+	// optimistic update and stays quiet.
+	ErrDeclined = errors.New("authorization declined")
+
+	// ErrUnavailable reports that this host has no elevation mechanism we can
+	// drive: no polkit on a Unix desktop, or an executable we decline to run as
+	// root. A caller falls back to telling the user which command to run.
+	ErrUnavailable = errors.New("no privilege elevation mechanism available")
+)
+
+// Run runs this executable with args under the platform's elevation mechanism
+// and waits for it to exit. A non-zero exit is returned as an error, so the
+// caller can treat a completed Run as the operation having succeeded.
+//
+// The args are the caller's own command line, so they cross no privilege
+// boundary: only a user who has just authenticated as an administrator can get
+// them run at all.
+func Run(ctx context.Context, args ...string) error {
+	self, err := trustedSelf()
+	if err != nil {
+		return err
+	}
+	return run(ctx, self, args)
+}
+
+// Available reports whether Run has a mechanism to use on this host, so a caller
+// can offer the prompt only when there is one and otherwise fall back to
+// guidance the user can act on. It answers from what is installed, not from what
+// the user is allowed to do: an administrator's password may still be required
+// and may still not be given, which is ErrDeclined from Run.
+func Available() bool {
+	if _, err := trustedSelf(); err != nil {
+		// Worth a line: this is also what a build run from a group-writable
+		// directory hits, and there is nothing in the UI to say why the offer is
+		// missing.
+		log.Debugf("not offering privilege elevation: %v", err)
+		return false
+	}
+	return mechanismAvailable()
+}

client/internal/elevate/output.go

@@ -0,0 +1,18 @@
+package elevate
+
+import "strings"
+
+// noOutput stands in for a process that said nothing, so that a report of what it
+// said still reads as a sentence.
+const noOutput = "no output"
+
+func firstLine(s string) string {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return noOutput
+	}
+	if i := strings.IndexByte(s, '\n'); i >= 0 {
+		return s[:i]
+	}
+	return s
+}

client/internal/elevate/output_test.go

@@ -0,0 +1,21 @@
+package elevate
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestFirstLine(t *testing.T) {
+	tests := []struct{ in, want string }{
+		{in: "", want: noOutput},
+		{in: "  \n ", want: noOutput},
+		{in: "one line", want: "one line"},
+		{in: "first\nsecond", want: "first"},
+		{in: "\nsecond\n", want: "second"},
+	}
+
+	for _, tt := range tests {
+		assert.Equal(t, tt.want, firstLine(tt.in), "input %q", tt.in)
+	}
+}

client/internal/elevate/run_darwin.go

@@ -0,0 +1,359 @@
+package elevate
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"runtime"
+	"strings"
+	"sync"
+	"syscall"
+	"unsafe"
+
+	"github.com/ebitengine/purego"
+	log "github.com/sirupsen/logrus"
+)
+
+// Authorization Services, reached through purego rather than cgo so the released
+// binaries keep building with CGO_ENABLED=0.
+//
+// The prompt belongs to this process, which is what makes it carry the
+// application's name and our own explanation. Going through osascript instead puts
+// the very same trampoline behind a dialog attributed to osascript, and means
+// handing a shell a command line to re-parse.
+//
+// # On AuthorizationExecuteWithPrivileges
+//
+// It is deprecated, and Apple's guidance (Quinn, "BSD Privilege Escalation on
+// macOS", developer.apple.com/forums/thread/708765) is "while it still works, it's
+// been deprecated for many years. Do not use it in a widely distributed product."
+// It is used here anyway, knowingly, because the alternatives Apple offers are for
+// *obtaining* ongoing privileges — an installer package, SMAppService, SMJobBless —
+// and NetBird already has what they would install: a launchd daemon running as
+// root. What is missing is only a way for an unprivileged client to ask it to act.
+//
+// The way to that without a deprecated call is to authorize the client instead of
+// elevating one: the app takes the right with AuthorizationCreate, passes the
+// AuthorizationExternalForm to the daemon, and the daemon checks it with
+// AuthorizationCopyRights before acting — none of which is deprecated. It is the
+// better design and it is where this should end up. It also means the daemon
+// accepting an authorization over its control socket, which is a new way to be
+// asked for privileged work and wants reviewing as such, so it is deliberately not
+// bundled in with the rest of this.
+//
+// Until then, three things keep the deprecation from being a trap. Every symbol is
+// resolved with an error rather than a panic, so a macOS that has dropped this
+// function leaves the app offering the user a command instead of crashing on the
+// way to a prompt. A failure to run the tool is reported as ErrUnavailable, so the
+// fallback is the same one an agent-less Linux session gets. And the whole path
+// runs under guard, which turns a panic out of the FFI layer into that same
+// fallback.
+//
+// The trampoline passes on the environment it was given, so what it starts as root
+// must be an executable this user's peers cannot influence: that is what
+// trustedSelf refuses, and what signing the binary settles for the loader.
+
+const (
+	securityFramework = "/System/Library/Frameworks/Security.framework/Security"
+	libSystem         = "/usr/lib/libSystem.B.dylib"
+
+	// trampoline is what the framework hands the tool to. Present on every macOS,
+	// and worth confirming before offering a prompt rather than mid-prompt.
+	trampoline = "/usr/libexec/security_authtrampoline"
+)
+
+// rightExecute is the right an administrator holds, and what
+// AuthorizationExecuteWithPrivileges requires of us.
+const rightExecute = "system.privilege.admin"
+
+// promptKey is kAuthorizationEnvironmentPrompt, which puts a sentence of ours above
+// the system's in the dialog. It is about the change rather than the mechanism.
+const (
+	promptKey  = "prompt"
+	promptText = "NetBird needs to change a setting that grants SSH access to this computer."
+)
+
+// OSStatus values from SecBase.h that mean something to us; anything else is
+// reported as it comes.
+const (
+	errAuthorizationSuccess               = 0
+	errAuthorizationDenied                = -60005
+	errAuthorizationCanceled              = -60006
+	errAuthorizationInteractionNotAllowed = -60007
+	errAuthorizationToolExecuteFailure    = -60031
+	errAuthorizationToolEnvironmentError  = -60032
+)
+
+// AuthorizationFlags from Authorization.h.
+const (
+	flagDefaults           = 0
+	flagInteractionAllowed = 1 << 0
+	flagExtendRights       = 1 << 1
+	flagDestroyRights      = 1 << 3
+	flagPreAuthorize       = 1 << 4
+)
+
+// authorizationItem mirrors AuthorizationItem: a name, and a value the name gives
+// meaning to. 32 bytes on both amd64 and arm64.
+type authorizationItem struct {
+	name        *byte
+	valueLength uintptr
+	value       unsafe.Pointer
+	// flags is reserved by the API and always zero. Declared because the layout
+	// is the contract: without it the struct is 24 bytes where C reads 32.
+	flags uint32 //nolint:unused // part of the C layout
+}
+
+// authorizationItemSet mirrors AuthorizationItemSet, which serves as both an
+// AuthorizationRights and an AuthorizationEnvironment.
+type authorizationItemSet struct {
+	count uint32
+	items *authorizationItem
+}
+
+var (
+	authorizationCreate                func(rights, environment *authorizationItemSet, flags uint32, authorization *uintptr) int32
+	authorizationExecuteWithPrivileges func(authorization uintptr, pathToTool string, options uint32, arguments *uintptr, communicationsPipe *uintptr) int32
+	authorizationFree                  func(authorization uintptr, flags uint32) int32
+	fileno                             func(stream uintptr) int32
+	fclose                             func(stream uintptr) int32
+
+	loadOnce sync.Once
+	loadErr  error
+)
+
+// load resolves the functions once. A framework that cannot be opened, or a symbol
+// that is no longer there, leaves the host without a mechanism rather than taking
+// the process down with it: see the note on deprecation above.
+func load() error {
+	loadOnce.Do(func() { loadErr = guard("loading Security.framework", resolve) })
+	return loadErr
+}
+
+// guard turns a panic out of the FFI layer into an error, so an API that has
+// changed under us costs the user a prompt rather than the window they were
+// clicking in. purego panics on a signature it cannot map, and this is the one
+// place in the client that calls a deprecated system function.
+//
+// It catches Go panics, which is what purego raises. A fault inside the framework
+// itself is not a panic and not recoverable; the layout the tests pin down is what
+// stands between us and that.
+func guard(what string, fn func() error) (err error) {
+	defer func() {
+		r := recover()
+		if r == nil {
+			return
+		}
+		log.Errorf("%s panicked: %v", what, r)
+		err = fmt.Errorf("%w: %s: %v", ErrUnavailable, what, r)
+	}()
+	return fn()
+}
+
+func resolve() error {
+	security, err := purego.Dlopen(securityFramework, purego.RTLD_LAZY|purego.RTLD_GLOBAL)
+	if err != nil {
+		return fmt.Errorf("open %s: %w", securityFramework, err)
+	}
+	system, err := purego.Dlopen(libSystem, purego.RTLD_LAZY|purego.RTLD_GLOBAL)
+	if err != nil {
+		return fmt.Errorf("open %s: %w", libSystem, err)
+	}
+
+	// purego.RegisterLibFunc panics on a symbol it cannot find, which is not how a
+	// deprecated function's disappearance should reach the user.
+	for _, fn := range []struct {
+		ptr    any
+		handle uintptr
+		name   string
+	}{
+		{&authorizationCreate, security, "AuthorizationCreate"},
+		{&authorizationExecuteWithPrivileges, security, "AuthorizationExecuteWithPrivileges"},
+		{&authorizationFree, security, "AuthorizationFree"},
+		{&fileno, system, "fileno"},
+		{&fclose, system, "fclose"},
+	} {
+		symbol, err := purego.Dlsym(fn.handle, fn.name)
+		if err != nil {
+			return fmt.Errorf("resolve %s: %w", fn.name, err)
+		}
+		if symbol == 0 {
+			return fmt.Errorf("resolve %s: not present on this system", fn.name)
+		}
+		purego.RegisterFunc(fn.ptr, symbol)
+	}
+	return nil
+}
+
+// run asks the system to run self as root: first for the right, which is what puts
+// up the authentication dialog and collects the password or takes the Touch ID,
+// then for the tool. The credentials go to the system's authorization trampoline
+// and never to us.
+//
+// The context bounds only our own waiting; the dialog belongs to the system and
+// closes when the user answers it.
+func run(ctx context.Context, self string, args []string) error {
+	if err := load(); err != nil {
+		return fmt.Errorf("%w: %v", ErrUnavailable, err)
+	}
+
+	return guard("asking for privileges", func() error {
+		authorization, err := authorize()
+		if err != nil {
+			return err
+		}
+		defer authorizationFree(authorization, flagDestroyRights)
+
+		return execute(ctx, authorization, self, args)
+	})
+}
+
+func mechanismAvailable() bool {
+	if err := load(); err != nil {
+		return false
+	}
+	info, err := os.Stat(trampoline)
+	return err == nil && !info.IsDir()
+}
+
+// authorize obtains the right, prompting for it. A dismissed dialog comes back as
+// errAuthorizationCanceled and a password given up on as errAuthorizationDenied;
+// both are the user's answer rather than a failure.
+func authorize() (uintptr, error) {
+	var pinner runtime.Pinner
+	defer pinner.Unpin()
+
+	rights := itemSet(&pinner, authorizationItem{name: cString(&pinner, rightExecute)})
+	environment := itemSet(&pinner, promptItem(&pinner))
+
+	var authorization uintptr
+	status := authorizationCreate(rights, environment,
+		flagDefaults|flagInteractionAllowed|flagPreAuthorize|flagExtendRights, &authorization)
+
+	switch status {
+	case errAuthorizationSuccess:
+		return authorization, nil
+	case errAuthorizationCanceled, errAuthorizationDenied:
+		return 0, ErrDeclined
+	case errAuthorizationInteractionNotAllowed:
+		// Nowhere to put a dialog, so there is nobody to ask: a launch daemon, or
+		// a session with no window server.
+		return 0, fmt.Errorf("%w: this session cannot show an authorization prompt", ErrUnavailable)
+	default:
+		return 0, fmt.Errorf("request %s: OSStatus %d", rightExecute, status)
+	}
+}
+
+// execute runs the tool with the right in hand and waits for it by reading the pipe
+// it is given until the tool closes it.
+//
+// AuthorizationExecuteWithPrivileges reports no exit status and does not say what
+// process it started, which is why the one-shot says so itself: what it prints is
+// the only evidence that the change was applied.
+func execute(ctx context.Context, authorization uintptr, self string, args []string) error {
+	var pinner runtime.Pinner
+	defer pinner.Unpin()
+
+	argv := make([]uintptr, 0, len(args)+1)
+	for _, arg := range args {
+		argv = append(argv, uintptr(unsafe.Pointer(cString(&pinner, arg))))
+	}
+	argv = append(argv, 0)
+	pinner.Pin(&argv[0])
+
+	var pipe uintptr
+	status := authorizationExecuteWithPrivileges(authorization, self, flagDefaults, &argv[0], &pipe)
+	switch status {
+	case errAuthorizationSuccess:
+	case errAuthorizationCanceled:
+		return ErrDeclined
+	case errAuthorizationToolExecuteFailure, errAuthorizationToolEnvironmentError:
+		// The right was granted and the tool still did not start. Nothing the user
+		// can do about it from here, so point them at the command instead.
+		return fmt.Errorf("%w: the system would not run %s elevated (OSStatus %d)", ErrUnavailable, self, status)
+	default:
+		return fmt.Errorf("run %s elevated: OSStatus %d", self, status)
+	}
+
+	out, err := readPipe(ctx, pipe)
+	if err != nil {
+		return err
+	}
+	return checkApplied(out)
+}
+
+// checkApplied reads the one-shot's report, which stands in for the exit status
+// there is no way to ask for here. A run that said nothing did not apply the
+// change, whatever else went on.
+func checkApplied(out string) error {
+	if !strings.Contains(out, AppliedMarker) {
+		return fmt.Errorf("elevated netbird did not report the change as applied: %s", firstLine(out))
+	}
+	return nil
+}
+
+// readPipe drains the tool's output, which ends when the tool exits and is
+// therefore also how we wait for it.
+func readPipe(ctx context.Context, pipe uintptr) (string, error) {
+	if pipe == 0 {
+		return "", nil
+	}
+	defer fclose(pipe)
+
+	fd := int(fileno(pipe))
+	if fd < 0 {
+		return "", nil
+	}
+
+	var out strings.Builder
+	buf := make([]byte, 4096)
+	for {
+		if err := ctx.Err(); err != nil {
+			return out.String(), err
+		}
+		n, err := syscall.Read(fd, buf)
+		if n > 0 {
+			out.Write(buf[:n])
+		}
+		switch {
+		case errors.Is(err, syscall.EINTR):
+			// A signal landed mid-read, which says nothing about the tool.
+			continue
+		case err != nil:
+			log.Debugf("read the elevated process's output: %v", err)
+			return out.String(), nil
+		case n <= 0:
+			// End of file: the tool closed the pipe, which is how it exiting
+			// reaches us.
+			return out.String(), nil
+		}
+	}
+}
+
+// itemSet builds an AuthorizationItemSet over items, pinned for the call.
+func itemSet(pinner *runtime.Pinner, items ...authorizationItem) *authorizationItemSet {
+	pinner.Pin(&items[0])
+	set := &authorizationItemSet{count: uint32(len(items)), items: &items[0]}
+	pinner.Pin(set)
+	return set
+}
+
+// promptItem is the environment entry carrying our sentence for the dialog.
+func promptItem(pinner *runtime.Pinner) authorizationItem {
+	value := []byte(promptText)
+	pinner.Pin(&value[0])
+	return authorizationItem{
+		name:        cString(pinner, promptKey),
+		valueLength: uintptr(len(value)),
+		value:       unsafe.Pointer(&value[0]),
+	}
+}
+
+// cString returns a NUL-terminated copy of s, pinned so the C side may hold it for
+// the duration of the call.
+func cString(pinner *runtime.Pinner, s string) *byte {
+	b := append([]byte(s), 0)
+	pinner.Pin(&b[0])
+	return &b[0]
+}

client/internal/elevate/run_darwin_test.go

@@ -0,0 +1,111 @@
+package elevate
+
+import (
+	"errors"
+	"runtime"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// The framework has to load and the symbols have to resolve, or nothing else here
+// means anything.
+func TestSecurityFrameworkLoads(t *testing.T) {
+	require.NoError(t, load(), "Security.framework must open")
+
+	for name, fn := range map[string]any{
+		"AuthorizationCreate":                authorizationCreate,
+		"AuthorizationExecuteWithPrivileges": authorizationExecuteWithPrivileges,
+		"AuthorizationFree":                  authorizationFree,
+		"fileno":                             fileno,
+		"fclose":                             fclose,
+	} {
+		assert.NotNil(t, fn, "%s must resolve", name)
+	}
+}
+
+// A request with no interaction allowed exercises the whole call — the rights and
+// environment structs, and the OSStatus that comes back — without a dialog anybody
+// has to answer. What the system decides is its business; that it decides at all is
+// what this asserts.
+func TestAuthorizationCreateWithoutInteraction(t *testing.T) {
+	if err := load(); err != nil {
+		t.Skipf("Security.framework did not open: %v", err)
+	}
+
+	var pinner runtime.Pinner
+	defer pinner.Unpin()
+
+	rights := itemSet(&pinner, authorizationItem{name: cString(&pinner, rightExecute)})
+	environment := itemSet(&pinner, promptItem(&pinner))
+	require.EqualValues(t, 1, rights.count, "the rights struct layout must match the C one")
+
+	var authorization uintptr
+	status := authorizationCreate(rights, environment, flagDefaults|flagExtendRights, &authorization)
+
+	switch status {
+	case errAuthorizationSuccess:
+		// Credentials were already cached for this session.
+		authorizationFree(authorization, flagDestroyRights)
+	case errAuthorizationDenied, errAuthorizationInteractionNotAllowed:
+		// The expected answers when nobody may be asked.
+	default:
+		require.Failf(t, "unknown OSStatus", "AuthorizationCreate returned %d, want a status we recognise", status)
+	}
+}
+
+// Asking with a right nobody has must not be mistaken for a declined prompt: the
+// caller would report nothing at all.
+func TestAuthorizeUnknownRightIsNotDeclined(t *testing.T) {
+	if err := load(); err != nil {
+		t.Skipf("Security.framework did not open: %v", err)
+	}
+
+	var pinner runtime.Pinner
+	defer pinner.Unpin()
+
+	rights := itemSet(&pinner, authorizationItem{name: cString(&pinner, "io.netbird.right.that.does.not.exist")})
+
+	var authorization uintptr
+	status := authorizationCreate(rights, nil, flagDefaults|flagExtendRights, &authorization)
+	if status == errAuthorizationSuccess {
+		authorizationFree(authorization, flagDestroyRights)
+	}
+	assert.NotEqual(t, int32(errAuthorizationSuccess), status, "a right that does not exist must not be granted")
+}
+
+func TestMechanismAvailable(t *testing.T) {
+	assert.True(t, mechanismAvailable(), "the trampoline exists on every macOS")
+}
+
+// The one-shot's report is what stands in for an exit status here, so a run that
+// says nothing must not read as success.
+func TestCheckApplied(t *testing.T) {
+	require.NoError(t, checkApplied(AppliedMarker+"\n"), "the report the one-shot prints")
+	require.NoError(t, checkApplied("some warning\n"+AppliedMarker+"\n"), "the report after other output")
+
+	assert.Error(t, checkApplied(""), "a run that printed nothing did not apply the change")
+	assert.Error(t, checkApplied("dyld: library not loaded\n"), "output that is not the report")
+}
+
+// A panic out of the FFI layer has to reach the caller as "no mechanism", which is
+// the outcome that offers the user the command instead of taking the window down.
+func TestGuardTurnsAPanicIntoUnavailable(t *testing.T) {
+	err := guard("pretending to call something", func() error {
+		panic("purego: signature it cannot map")
+	})
+
+	require.ErrorIs(t, err, ErrUnavailable, "a panic must read as a missing mechanism")
+	assert.Contains(t, err.Error(), "pretending to call something", "what panicked")
+}
+
+// guard wraps every darwin path, so what a caller switches on has to survive it.
+func TestGuardPassesErrorsThrough(t *testing.T) {
+	sentinel := errors.New("the call itself failed")
+	assert.ErrorIs(t, guard("calling", func() error { return sentinel }), sentinel,
+		"the error it was given")
+	assert.ErrorIs(t, guard("calling", func() error { return ErrDeclined }), ErrDeclined,
+		"a declined prompt stays declined")
+	assert.NoError(t, guard("calling", func() error { return nil }), "a call that worked")
+}

client/internal/elevate/run_unix.go

@@ -0,0 +1,117 @@
+//go:build linux
+
+package elevate
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"strings"
+)
+
+// pkexec exit codes that are about the authorization rather than about the program
+// we asked it to run. The manual page reserves both.
+const (
+	// exitDismissed is returned when the user dismissed the authentication
+	// dialog.
+	exitDismissed = 126
+	// exitNotAuthorized is returned when the authorization was not obtained. That
+	// covers the user saying no as well as pkexec having had nobody to ask: see
+	// noAgentMarkers.
+	exitNotAuthorized = 127
+)
+
+// exitNotAuthorized covers three different endings that only pkexec's own words
+// tell apart, so they are matched here. Read with LC_ALL=C so the words are the
+// ones written below.
+//
+// refusedMarker is a refusal: the user said no, gave up on the password, or holds
+// an account that may not elevate at all.
+const refusedMarker = "Not authorized"
+
+// noAgentMarkers say pkexec had no way to ask: no agent registered for the
+// session, and no controlling terminal for the textual agent it falls back to.
+var noAgentMarkers = []string{"authentication agent", "controlling terminal"}
+
+// run asks polkit to run self as root. pkexec hands the request to the session's
+// polkit agent, which is what prompts and what collects any password; we see only
+// its verdict.
+//
+// The environment is otherwise deliberately not passed through: pkexec clears it
+// bar a small allowlist, and the one-shot needs nothing from it.
+func run(ctx context.Context, self string, args []string) error {
+	pkexec, err := exec.LookPath("pkexec")
+	if err != nil {
+		return fmt.Errorf("%w: pkexec is not installed", ErrUnavailable)
+	}
+
+	cmd := exec.CommandContext(ctx, pkexec, append([]string{self}, args...)...)
+	// C locale so pkexec's own diagnostics are the ones noAgentMarkers knows.
+	cmd.Env = append(os.Environ(), "LC_ALL=C")
+	var stderr strings.Builder
+	cmd.Stderr = &stderr
+	// The one-shot reports itself on stdout for macOS's sake, where there is no
+	// exit status to read. Here there is one, so that line is noise.
+	cmd.Stdout = io.Discard
+
+	err = cmd.Run()
+	if err == nil {
+		return nil
+	}
+
+	var exitErr *exec.ExitError
+	if !errors.As(err, &exitErr) {
+		return fmt.Errorf("run pkexec: %w", err)
+	}
+
+	// Matched against everything pkexec said, reported as one line: a complaint
+	// that is not the first thing printed still has to be recognised, and reading
+	// it as a refusal would swallow it.
+	full := stderr.String()
+	out := firstLine(full)
+
+	switch exitErr.ExitCode() {
+	case exitDismissed:
+		return ErrDeclined
+	case exitNotAuthorized:
+		return notAuthorized(full, out)
+	default:
+		return fmt.Errorf("elevated netbird exited with %d: %s", exitErr.ExitCode(), out)
+	}
+}
+
+// notAuthorized sorts out the three endings pkexec reports as exitNotAuthorized.
+//
+// It also returns that code when the authorization succeeded and it then could
+// not run the program, so a refusal has to be recognised rather than assumed:
+// reading every one of these as "the user said no" would revert the control in
+// silence on a host where elevation is broken.
+func notAuthorized(full, out string) error {
+	switch {
+	case hasAny(full, noAgentMarkers):
+		return fmt.Errorf("%w: polkit had no way to ask: %s", ErrUnavailable, out)
+	case out == noOutput, strings.Contains(full, refusedMarker):
+		// The user said no, which needs no message; that an account barred from
+		// elevating altogether lands here too is why the reason is kept.
+		return fmt.Errorf("%w: %s", ErrDeclined, out)
+	default:
+		return fmt.Errorf("pkexec could not run elevated netbird: %s", out)
+	}
+}
+
+func hasAny(s string, markers []string) bool {
+	for _, marker := range markers {
+		if strings.Contains(s, marker) {
+			return true
+		}
+	}
+	return false
+}
+
+func mechanismAvailable() bool {
+	_, err := exec.LookPath("pkexec")
+	return err == nil
+}

client/internal/elevate/run_unix_test.go

@@ -0,0 +1,110 @@
+//go:build linux
+
+package elevate
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// fakePkexec puts a pkexec on PATH that exits with the given code, so the
+// mapping from polkit's exit codes onto our errors can be exercised without a
+// polkit agent.
+func fakePkexec(t *testing.T, exitCode int, stderr string) {
+	t.Helper()
+
+	dir := t.TempDir()
+	script := fmt.Sprintf("#!/bin/sh\necho %s >&2\nexit %d\n", shellQuote(stderr), exitCode)
+	require.NoError(t, os.WriteFile(filepath.Join(dir, "pkexec"), []byte(script), 0o700), "write the fake pkexec")
+	t.Setenv("PATH", dir)
+}
+
+func shellQuote(s string) string {
+	return "'" + strings.ReplaceAll(s, "'", `'\''`) + "'"
+}
+
+func TestRunMapsPkexecExitCodes(t *testing.T) {
+	tests := []struct {
+		name     string
+		exitCode int
+		stderr   string
+		wantErr  error
+	}{
+		{name: "applied", exitCode: 0},
+		{
+			name:     "dialog dismissed",
+			exitCode: exitDismissed,
+			stderr:   "Error executing command as another user: Request dismissed",
+			wantErr:  ErrDeclined,
+		},
+		{
+			// What a graphical agent reports for a cancelled prompt. Not a
+			// failure: the user was asked and answered.
+			name:     "prompt cancelled",
+			exitCode: exitNotAuthorized,
+			stderr:   "Error executing command as another user: Not authorized",
+			wantErr:  ErrDeclined,
+		},
+		{
+			// The same status, but pkexec never got to ask anybody.
+			name:     "no agent and no terminal to fall back on",
+			exitCode: exitNotAuthorized,
+			stderr:   "Error creating textual authentication agent: Error opening current controlling terminal for the process (`/dev/tty'): No such device or address",
+			wantErr:  ErrUnavailable,
+		},
+		{
+			// And the same status again once the authorization succeeded and
+			// pkexec could not run what it had been authorized to run. Reading
+			// that as a refusal would revert the control in silence on a host
+			// where elevation is broken.
+			name:     "authorized but not runnable",
+			exitCode: exitNotAuthorized,
+			stderr:   "Error executing command as another user: No such file or directory",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			fakePkexec(t, tt.exitCode, tt.stderr)
+
+			err := run(context.Background(), "/nonexistent/netbird-ui", []string{"--flag"})
+			switch {
+			case tt.wantErr != nil:
+				require.ErrorIs(t, err, tt.wantErr, "exit %d said %q", tt.exitCode, tt.stderr)
+			case tt.exitCode == 0:
+				require.NoError(t, err, "a pkexec that exited cleanly applied the change")
+			default:
+				require.Error(t, err, "exit %d said %q", tt.exitCode, tt.stderr)
+				assert.NotErrorIs(t, err, ErrDeclined, "not the user's answer")
+				assert.NotErrorIs(t, err, ErrUnavailable, "not a missing mechanism")
+			}
+		})
+	}
+}
+
+// An exit code that is not polkit's is the one-shot's own failure, and has to
+// stay distinguishable from a declined prompt: the caller reports it.
+func TestRunReportsOneShotFailure(t *testing.T) {
+	fakePkexec(t, 3, "the one-shot said no")
+
+	err := run(context.Background(), "/nonexistent/netbird-ui", nil)
+
+	require.Error(t, err, "a one-shot that failed is not a prompt that was answered")
+	assert.NotErrorIs(t, err, ErrDeclined, "not the user's answer")
+	assert.NotErrorIs(t, err, ErrUnavailable, "not a missing mechanism")
+}
+
+func TestRunWithoutPkexecIsUnavailable(t *testing.T) {
+	t.Setenv("PATH", t.TempDir())
+
+	err := run(context.Background(), "/nonexistent/netbird-ui", nil)
+	require.ErrorIs(t, err, ErrUnavailable, "no pkexec means no mechanism")
+	assert.False(t, mechanismAvailable(), "mechanismAvailable without pkexec on PATH")
+}

client/internal/elevate/run_unsupported.go

@@ -0,0 +1,19 @@
+//go:build !windows && !darwin && !linux
+
+package elevate
+
+import "context"
+
+// run reports that this platform has no elevation prompt to drive.
+//
+// The desktop app is the only caller and is not built for any of these: mobile
+// and WASM have no local user to ask, and the FreeBSD client ships without a UI.
+// pkexec would be the mechanism there, and run_unix.go is what to widen if that
+// changes.
+func run(context.Context, string, []string) error {
+	return ErrUnavailable
+}
+
+func mechanismAvailable() bool {
+	return false
+}

client/internal/elevate/run_windows.go

@@ -0,0 +1,187 @@
+package elevate
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"runtime"
+	"unsafe"
+
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/sys/windows"
+)
+
+const (
+	// seeMaskNoCloseProcess keeps the started process's handle open in
+	// hProcess so we can wait for it.
+	seeMaskNoCloseProcess = 0x00000040
+	// seeMaskNoAsync makes ShellExecuteExW finish its work before returning,
+	// which it must when the calling thread does not pump messages.
+	seeMaskNoAsync = 0x00000100
+	// seeMaskFlagNoUI suppresses the shell's own error dialogs; the UAC consent
+	// dialog is not one of them and still appears.
+	seeMaskFlagNoUI = 0x00000400
+
+	// swHide: the one-shot has no window to show.
+	swHide = 0
+)
+
+// shellExecuteInfoW mirrors SHELLEXECUTEINFOW. The field order and Go's own
+// padding match the C layout on both 386 and amd64.
+type shellExecuteInfoW struct {
+	cbSize         uint32
+	fMask          uint32
+	hwnd           windows.HWND
+	lpVerb         *uint16
+	lpFile         *uint16
+	lpParameters   *uint16
+	lpDirectory    *uint16
+	nShow          int32
+	hInstApp       windows.Handle
+	lpIDList       uintptr
+	lpClass        *uint16
+	hkeyClass      windows.Handle
+	dwHotKey       uint32
+	hIconOrMonitor windows.Handle
+	hProcess       windows.Handle
+}
+
+var (
+	shell32            = windows.NewLazySystemDLL("shell32.dll")
+	procShellExecuteEx = shell32.NewProc("ShellExecuteExW")
+)
+
+// run starts self elevated with the "runas" verb, which is what raises the UAC
+// consent dialog, and waits for it to finish. Windows decides whether consent is
+// enough or an administrator's credentials are needed, and collects them itself.
+func run(ctx context.Context, self string, args []string) error {
+	verb, err := windows.UTF16PtrFromString("runas")
+	if err != nil {
+		return fmt.Errorf("encode verb: %w", err)
+	}
+	file, err := windows.UTF16PtrFromString(self)
+	if err != nil {
+		return fmt.Errorf("encode %s: %w", self, err)
+	}
+	params, err := windows.UTF16PtrFromString(windows.ComposeCommandLine(args))
+	if err != nil {
+		return fmt.Errorf("encode arguments: %w", err)
+	}
+
+	info := shellExecuteInfoW{
+		fMask:        seeMaskNoCloseProcess | seeMaskNoAsync | seeMaskFlagNoUI,
+		hwnd:         ownerWindow(),
+		lpVerb:       verb,
+		lpFile:       file,
+		lpParameters: params,
+		nShow:        swHide,
+	}
+	info.cbSize = uint32(unsafe.Sizeof(info))
+
+	process, err := shellExecute(&info)
+	if err != nil {
+		return err
+	}
+	defer func() {
+		if err := windows.CloseHandle(process); err != nil {
+			log.Debugf("close elevated process handle: %v", err)
+		}
+	}()
+
+	return waitForProcess(ctx, process)
+}
+
+// shellExecute performs the call itself. ShellExecuteExW wants COM initialised on
+// the calling thread, so the goroutine is pinned to one for the duration and COM
+// is set up on it; an "already initialised, different mode" answer is fine,
+// because then somebody else has done it for us.
+func shellExecute(info *shellExecuteInfoW) (windows.Handle, error) {
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+
+	switch err := windows.CoInitializeEx(0, windows.COINIT_APARTMENTTHREADED); {
+	case err == nil, isHResult(err, windows.S_FALSE):
+		// Ours, or already initialised in the same mode: either way this call
+		// counts and has to be balanced.
+		defer windows.CoUninitialize()
+	case isHResult(err, windows.RPC_E_CHANGED_MODE):
+		// The thread is already in the other apartment model. ShellExecuteExW
+		// works there too, and there is nothing of ours to balance.
+	default:
+		return 0, fmt.Errorf("initialise COM: %w", err)
+	}
+
+	ret, _, lastErr := procShellExecuteEx.Call(uintptr(unsafe.Pointer(info)))
+	if ret != 0 {
+		return info.hProcess, nil
+	}
+
+	if errors.Is(lastErr, windows.ERROR_CANCELLED) {
+		return 0, ErrDeclined
+	}
+	return 0, fmt.Errorf("run elevated: %w", lastErr)
+}
+
+// ownerWindow returns this process's foreground window, and 0 when the window in
+// front belongs to somebody else or cannot be attributed. ShellExecuteExW takes it
+// as the parent for the UI it raises, which is what keeps the consent dialog in
+// front of the window the user was just clicking in instead of behind it. It is
+// also what a remote-desktop session needs to place the dialog at all when the
+// secure desktop is switched off.
+func ownerWindow() windows.HWND {
+	hwnd := windows.GetForegroundWindow()
+	if hwnd == 0 {
+		return 0
+	}
+
+	var pid uint32
+	if _, err := windows.GetWindowThreadProcessId(hwnd, &pid); err != nil {
+		log.Debugf("cannot attribute the foreground window, raising the prompt without an owner: %v", err)
+		return 0
+	}
+	if pid != windows.GetCurrentProcessId() {
+		return 0
+	}
+	return hwnd
+}
+
+// isHResult reports whether err carries the given HRESULT. CoInitializeEx
+// returns its HRESULT as an Errno, so the comparison is on the raw value.
+func isHResult(err error, hresult windows.Handle) bool {
+	var errno windows.Errno
+	return errors.As(err, &errno) && uintptr(errno) == uintptr(hresult)
+}
+
+func waitForProcess(ctx context.Context, process windows.Handle) error {
+	// The wait is interruptible so a cancelled context stops us waiting on a
+	// consent dialog nobody is going to answer. The elevated process is not
+	// ours to kill, and it either applies the change or does not.
+	for {
+		event, err := windows.WaitForSingleObject(process, 250)
+		if err != nil {
+			return fmt.Errorf("wait for the elevated process: %w", err)
+		}
+		if event == uint32(windows.WAIT_OBJECT_0) {
+			break
+		}
+		if err := ctx.Err(); err != nil {
+			return err
+		}
+	}
+
+	var code uint32
+	if err := windows.GetExitCodeProcess(process, &code); err != nil {
+		return fmt.Errorf("read the elevated process's exit code: %w", err)
+	}
+	if code != 0 {
+		return fmt.Errorf("elevated netbird exited with %d", code)
+	}
+	return nil
+}
+
+// mechanismAvailable is true on Windows: UAC prompts for consent when the user
+// is an administrator and for an administrator's credentials when they are not,
+// so there is always something to ask.
+func mechanismAvailable() bool {
+	return true
+}

client/internal/elevate/trusted.go

@@ -0,0 +1,40 @@
+package elevate
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+)
+
+// trustedSelf returns the path of this executable, provided it is one we are
+// willing to have run as root.
+//
+// The check is what keeps elevation from becoming a way to launder someone
+// else's code into a root process: the user consents to NetBird being elevated,
+// having been shown NetBird's name, so what runs must be the file NetBird was
+// installed as and not something a third party could have swapped for it. An
+// executable only its owner can write is that; anything wider is refused, and
+// the caller falls back to showing the command instead.
+//
+// The owner writing to their own executable is not part of that threat: code
+// running as the user can already prompt them for anything, and could just as
+// well ask them to run the command by hand. What matters is that no *other*
+// unprivileged account can reach it.
+func trustedSelf() (string, error) {
+	exe, err := os.Executable()
+	if err != nil {
+		return "", fmt.Errorf("locate this executable: %w", err)
+	}
+
+	// Resolve symlinks so the checks below apply to the file that would actually
+	// be executed, not to a link somebody else may control.
+	resolved, err := filepath.EvalSymlinks(exe)
+	if err != nil {
+		return "", fmt.Errorf("resolve %s: %w", exe, err)
+	}
+
+	if err := checkOnlyOwnerWritable(resolved); err != nil {
+		return "", fmt.Errorf("%w: %s cannot be trusted to run as root: %w", ErrUnavailable, resolved, err)
+	}
+	return resolved, nil
+}

client/internal/elevate/trusted_group_darwin.go

@@ -0,0 +1,10 @@
+package elevate
+
+// adminWriteGIDs are the groups whose write access to an executable does not
+// widen who could authorize elevating it.
+//
+// macOS installs applications as root:admin, mode 0775, /Applications included,
+// so requiring owner-only write would reject every normal install. Group admin
+// (gid 80) is exactly the set of accounts that can answer the authentication
+// dialog, so its write access grants nothing the prompt would not.
+var adminWriteGIDs = []uint32{0, 80}

client/internal/elevate/trusted_group_unix.go

@@ -0,0 +1,9 @@
+//go:build !windows && !darwin
+
+package elevate
+
+// adminWriteGIDs are the groups whose write access to an executable does not
+// widen who could authorize elevating it. Only root's own group qualifies here:
+// a distribution installs into root-owned directories, and there is no
+// system-wide administrators group that both writes them and answers polkit.
+var adminWriteGIDs = []uint32{0}

client/internal/elevate/trusted_unix.go

@@ -0,0 +1,119 @@
+//go:build !windows
+
+package elevate
+
+import (
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"slices"
+	"strconv"
+	"syscall"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/getent"
+)
+
+// checkOnlyOwnerWritable reports an error unless path, and every directory leading
+// to it, is owned by either root or this user and writable by nobody who could not
+// already act as its owner. A writable directory is as good as a writable file,
+// since anything in it can be replaced, so the whole chain is checked.
+func checkOnlyOwnerWritable(path string) error {
+	self := uint32(os.Getuid())
+
+	for dir := path; ; dir = filepath.Dir(dir) {
+		info, err := os.Lstat(dir)
+		if err != nil {
+			return fmt.Errorf("stat %s: %w", dir, err)
+		}
+
+		stat, ok := info.Sys().(*syscall.Stat_t)
+		if !ok {
+			return errors.New("file ownership is unavailable on this platform")
+		}
+		if stat.Uid != 0 && stat.Uid != self {
+			return fmt.Errorf("%s is owned by uid %d, neither root nor this user", dir, stat.Uid)
+		}
+
+		if err := checkWriteBits(dir, info, stat.Uid, stat.Gid); err != nil {
+			return err
+		}
+
+		if parent := filepath.Dir(dir); parent == dir {
+			return nil
+		}
+	}
+}
+
+func checkWriteBits(path string, info os.FileInfo, uid, gid uint32) error {
+	// On a directory the sticky bit stands in for the write bits: whoever may
+	// write there still cannot replace an entry they do not own, which is the
+	// only thing that would matter to us. /tmp is the usual example.
+	sticky := info.IsDir() && info.Mode()&os.ModeSticky != 0
+
+	return writeBitsAllow(path, info.Mode().Perm(), sticky, groupWriteAllowed(uid, gid))
+}
+
+// writeBitsAllow decides on the permission bits alone, given whether the group's
+// write access has been vouched for.
+func writeBitsAllow(path string, perm os.FileMode, sticky, groupAllowed bool) error {
+	if sticky {
+		return nil
+	}
+	if perm&0o020 != 0 && !groupAllowed {
+		return fmt.Errorf("%s is writable by a group with members other than its owner (%v)", path, perm)
+	}
+	if perm&0o002 != 0 {
+		return fmt.Errorf("%s is world-writable (%v)", path, perm)
+	}
+	return nil
+}
+
+// groupWriteAllowed reports whether a group's write access to a file owned by uid
+// puts it in reach of anyone who could not already act as that owner.
+//
+// Two ways it does not. A group in adminWriteGIDs holds the accounts that can
+// answer the elevation prompt anyway. And a user private group is how Debian,
+// Ubuntu and Fedora ship: their umask of 002 makes a home directory and
+// everything built in it group-writable, so refusing that would refuse every
+// build not installed from a package.
+func groupWriteAllowed(uid, gid uint32) bool {
+	if slices.Contains(adminWriteGIDs, gid) {
+		return true
+	}
+
+	group, err := getent.LookupGroupID(strconv.FormatUint(uint64(gid), 10))
+	if err != nil {
+		log.Debugf("cannot look up group %d, treating it as shared: %v", gid, err)
+		return false
+	}
+	owner, err := getent.LookupUserID(strconv.FormatUint(uint64(uid), 10))
+	if err != nil {
+		log.Debugf("cannot look up uid %d, treating its group as shared: %v", uid, err)
+		return false
+	}
+
+	if group.Name != owner.Username {
+		return false
+	}
+	return !groupHasOtherMembers(group.Name, owner.Username)
+}
+
+// groupHasOtherMembers reports whether the group lists a member besides owner.
+//
+// Sharing the owner's name is what a user private group is recognised by, and it
+// says nothing about who is in it: a group that has since gained a member is
+// still named that way, and that member can write whatever the group can. So the
+// membership is read rather than assumed. A group whose members cannot be
+// listed, because no source on this host describes it, is treated as shared:
+// the name alone cannot vouch for who writes through it.
+func groupHasOtherMembers(name, owner string) bool {
+	members, err := getent.GroupMembers(name)
+	if err != nil {
+		log.Debugf("cannot list the members of group %q, treating it as shared: %v", name, err)
+		return true
+	}
+	return slices.ContainsFunc(members, func(member string) bool { return member != owner })
+}

client/internal/elevate/trusted_unix_test.go

@@ -0,0 +1,148 @@
+//go:build !windows
+
+package elevate
+
+import (
+	"os"
+	"os/user"
+	"path/filepath"
+	"strconv"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// ownerOnlyDir is t.TempDir() with the write bits tightened. testing creates its
+// numbered directory with 0777 minus the umask, so under the common 002 umask it
+// is group-writable and would fail the check under test on its own.
+func ownerOnlyDir(t *testing.T) string {
+	t.Helper()
+	dir := t.TempDir()
+	require.NoError(t, os.Chmod(dir, 0o755), "tighten the temporary directory")
+	return dir
+}
+
+// writeExecutable creates a plain executable file, the shape trustedSelf checks.
+func writeExecutable(t *testing.T, dir string) string {
+	t.Helper()
+	path := filepath.Join(dir, "netbird-ui")
+	require.NoError(t, os.WriteFile(path, []byte("#!/bin/sh\n"), 0o755), "write the executable")
+	require.NoError(t, os.Chmod(path, 0o755), "set the executable's mode")
+	return path
+}
+
+func TestCheckOnlyOwnerWritableAcceptsOwnerOnly(t *testing.T) {
+	err := checkOnlyOwnerWritable(writeExecutable(t, ownerOnlyDir(t)))
+	assert.NoError(t, err, "an owner-only writable executable is trustworthy")
+}
+
+func TestCheckOnlyOwnerWritableRejectsWorldWritableFile(t *testing.T) {
+	path := writeExecutable(t, ownerOnlyDir(t))
+	require.NoError(t, os.Chmod(path, 0o777), "make the executable world-writable")
+
+	assert.Error(t, checkOnlyOwnerWritable(path), "a world-writable executable must be refused")
+}
+
+// The permission policy on its own, without a filesystem to arrange: whether the
+// group has been vouched for is the only thing that makes group write acceptable.
+func TestWriteBitsAllow(t *testing.T) {
+	tests := []struct {
+		name         string
+		perm         os.FileMode
+		sticky       bool
+		groupAllowed bool
+		wantErr      bool
+	}{
+		{name: "owner only", perm: 0o755},
+		{name: "group write in a private group", perm: 0o775, groupAllowed: true},
+		{name: "group write in a shared group", perm: 0o775, wantErr: true},
+		{name: "world write", perm: 0o777, groupAllowed: true, wantErr: true},
+		{name: "world write on a sticky directory", perm: 0o777, sticky: true},
+		{name: "group write on a sticky directory", perm: 0o775, sticky: true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			err := writeBitsAllow("/path", tt.perm, tt.sticky, tt.groupAllowed)
+			if tt.wantErr {
+				assert.Error(t, err, "perm %v, sticky %v, group allowed %v", tt.perm, tt.sticky, tt.groupAllowed)
+				return
+			}
+			assert.NoError(t, err, "perm %v, sticky %v, group allowed %v", tt.perm, tt.sticky, tt.groupAllowed)
+		})
+	}
+}
+
+// A build under a home directory on a distribution with a 002 umask, which is what
+// a locally built or tarball-installed binary looks like. Its group has no members
+// but its owner, so it is as good as owner-only.
+//
+// Whether this host is such a distribution is read from the environment rather than
+// from groupWriteAllowed: asking the function under test whether to run would let
+// it skip its own coverage away if it regressed to refusing everything.
+func TestCheckOnlyOwnerWritableAcceptsOwnPrivateGroup(t *testing.T) {
+	requirePrivatePrimaryGroup(t)
+
+	dir := ownerOnlyDir(t)
+	path := writeExecutable(t, dir)
+	require.NoError(t, os.Chmod(dir, 0o775), "make the directory group-writable")
+	require.NoError(t, os.Chmod(path, 0o775), "make the executable group-writable")
+
+	err := checkOnlyOwnerWritable(path)
+	assert.NoError(t, err, "group write in the owner's own private group reaches nobody else")
+}
+
+// A group whose membership no source can answer for is treated as shared: the
+// private-group allowance must not stand on a name nobody can vouch for. The
+// membership listing itself lives in the getent package and is tested there.
+func TestGroupHasOtherMembersRejectsAnUnknownGroup(t *testing.T) {
+	assert.True(t, groupHasOtherMembers("nonexistent_group_xyzzy_12345", "vma"),
+		"a group no source describes")
+}
+
+// A writable directory is as good as a writable file: whoever can write the
+// directory can put a different binary at the same path.
+func TestCheckOnlyOwnerWritableRejectsWritableDirectory(t *testing.T) {
+	dir := filepath.Join(ownerOnlyDir(t), "bin")
+	require.NoError(t, os.Mkdir(dir, 0o755), "create the directory")
+	path := writeExecutable(t, dir)
+	require.NoError(t, os.Chmod(dir, 0o777), "make the directory world-writable")
+
+	assert.Error(t, checkOnlyOwnerWritable(path), "an executable in a world-writable directory must be refused")
+}
+
+// A sticky world-writable directory is exempt: the sticky bit is what stops one
+// user replacing another's entries. /tmp is why this matters.
+func TestCheckOnlyOwnerWritableAcceptsStickyDirectory(t *testing.T) {
+	dir := filepath.Join(ownerOnlyDir(t), "sticky")
+	require.NoError(t, os.Mkdir(dir, 0o755), "create the directory")
+	path := writeExecutable(t, dir)
+	require.NoError(t, os.Chmod(dir, 0o777|os.ModeSticky), "make the directory sticky and world-writable")
+
+	err := checkOnlyOwnerWritable(path)
+	assert.NoError(t, err, "the sticky bit stops another user replacing the executable")
+}
+
+func TestCheckOnlyOwnerWritableRejectsMissingFile(t *testing.T) {
+	err := checkOnlyOwnerWritable(filepath.Join(ownerOnlyDir(t), "absent"))
+	assert.Error(t, err, "an executable that is not there must be refused")
+}
+
+// requirePrivatePrimaryGroup skips unless this user's primary group is their own,
+// which is what the user-private-group allowance is about.
+func requirePrivatePrimaryGroup(t *testing.T) {
+	t.Helper()
+
+	self, err := user.Current()
+	require.NoError(t, err, "look up the test user")
+	group, err := user.LookupGroupId(strconv.Itoa(os.Getgid()))
+	require.NoError(t, err, "look up the test user's primary group")
+
+	if group.Name != self.Username {
+		t.Skipf("the test user's primary group is %q, not their own, so there is nothing to assert here", group.Name)
+	}
+	if groupHasOtherMembers(group.Name, self.Username) {
+		t.Skipf("group %q has other members, so it is not a private group", group.Name)
+	}
+}

client/internal/elevate/trusted_windows.go

@@ -0,0 +1,215 @@
+package elevate
+
+import (
+	"errors"
+	"fmt"
+	"path/filepath"
+	"slices"
+	"unsafe"
+
+	"golang.org/x/sys/windows"
+)
+
+const (
+	// fileDeleteChild is FILE_DELETE_CHILD, which x/sys does not define: the
+	// right to delete an entry of a directory without holding DELETE on it.
+	fileDeleteChild = 0x00000040
+
+	// accessAllowedCallbackACEType is an allow ACE with a condition appended to
+	// the ACCESS_ALLOWED_ACE layout, so its trustee is still at SidStart.
+	accessAllowedCallbackACEType = 0x9
+
+	// The allow ACE types that carry object GUIDs ahead of the trustee, so the
+	// SID is not at SidStart. They occur on directory-service objects rather
+	// than files, and are refused rather than skipped: see aceTrustee.
+	accessAllowedObjectACEType         = 0x5
+	accessAllowedCallbackObjectACEType = 0xB
+)
+
+// fileWriteAccess are the rights that let a trustee rewrite or replace a file,
+// or take it over and then do so.
+const fileWriteAccess = windows.FILE_WRITE_DATA | windows.FILE_APPEND_DATA |
+	windows.DELETE | windows.WRITE_DAC | windows.WRITE_OWNER |
+	windows.GENERIC_WRITE | windows.GENERIC_ALL
+
+// dirWriteAccess are the rights over a directory that let a trustee replace an
+// entry somebody else owns. Creating a new entry is not one of them, which is
+// what the Unix sticky bit says in one bit: the root of every volume grants
+// BUILTIN\Users the right to add directories under it, and that reaches nothing
+// already there.
+const dirWriteAccess = fileDeleteChild | windows.DELETE |
+	windows.WRITE_DAC | windows.WRITE_OWNER | windows.GENERIC_ALL
+
+// trustedInstallerSID owns much of what Windows itself installs. x/sys has no
+// well-known constant for it.
+const trustedInstallerSID = "S-1-5-80-956008885-3418522649-1831038044-1853292631-2271478464"
+
+// checkOnlyOwnerWritable reports an error unless path, and every directory
+// leading to it, is owned by an account that can elevate (or by this user) and
+// grants write access to nobody else. A writable directory is as good as a
+// writable file, since an entry in it can be replaced, so the whole chain is
+// checked.
+func checkOnlyOwnerWritable(path string) error {
+	owners, err := trustedOwners()
+	if err != nil {
+		return err
+	}
+	writers, err := trustedWriters(owners)
+	if err != nil {
+		return err
+	}
+
+	writeAccess := windows.ACCESS_MASK(fileWriteAccess)
+	for target := path; ; target = filepath.Dir(target) {
+		if err := checkSecurity(target, writeAccess, owners, writers); err != nil {
+			return err
+		}
+		if parent := filepath.Dir(target); parent == target {
+			return nil
+		}
+		writeAccess = dirWriteAccess
+	}
+}
+
+// trustedOwners are the accounts we accept as the owner of the executable and of
+// the directories above it: the ones that can already answer the UAC prompt,
+// plus this user, whose own executable is theirs to write. Code running as the
+// user could prompt them for anything anyway; what matters is that no *other*
+// unprivileged account can reach it.
+func trustedOwners() ([]*windows.SID, error) {
+	self, err := currentUserSID()
+	if err != nil {
+		return nil, err
+	}
+
+	owners := []*windows.SID{self}
+	for _, wellKnown := range []windows.WELL_KNOWN_SID_TYPE{
+		windows.WinLocalSystemSid,
+		windows.WinBuiltinAdministratorsSid,
+	} {
+		sid, err := windows.CreateWellKnownSid(wellKnown)
+		if err != nil {
+			return nil, fmt.Errorf("build well-known SID %d: %w", wellKnown, err)
+		}
+		owners = append(owners, sid)
+	}
+
+	installer, err := windows.StringToSid(trustedInstallerSID)
+	if err != nil {
+		return nil, fmt.Errorf("parse TrustedInstaller SID: %w", err)
+	}
+	return append(owners, installer), nil
+}
+
+// trustedWriters are the trustees whose write access does not widen who could
+// decide what runs behind the prompt. The owners, and CREATOR OWNER, which
+// resolves to the object's owner and is therefore already vetted.
+func trustedWriters(owners []*windows.SID) ([]*windows.SID, error) {
+	creatorOwner, err := windows.CreateWellKnownSid(windows.WinCreatorOwnerSid)
+	if err != nil {
+		return nil, fmt.Errorf("build the CREATOR OWNER SID: %w", err)
+	}
+	return append(slices.Clone(owners), creatorOwner), nil
+}
+
+func checkSecurity(path string, writeAccess windows.ACCESS_MASK, owners, writers []*windows.SID) error {
+	sd, err := windows.GetNamedSecurityInfo(path, windows.SE_FILE_OBJECT,
+		windows.OWNER_SECURITY_INFORMATION|windows.DACL_SECURITY_INFORMATION)
+	if err != nil {
+		return fmt.Errorf("read security descriptor of %s: %w", path, err)
+	}
+
+	owner, _, err := sd.Owner()
+	if err != nil {
+		return fmt.Errorf("read owner of %s: %w", path, err)
+	}
+	if !containsSID(owners, owner) {
+		return fmt.Errorf("%s is owned by %s, which is neither this user nor an account that can elevate", path, owner)
+	}
+
+	dacl, _, err := sd.DACL()
+	if err != nil {
+		return fmt.Errorf("read DACL of %s: %w", path, err)
+	}
+	// A NULL DACL grants everyone everything; only an absent security
+	// descriptor would have got us here without one, and neither is trustworthy.
+	if dacl == nil {
+		return fmt.Errorf("%s has no DACL, so it grants write access to everyone", path)
+	}
+
+	return checkDACL(path, dacl, writeAccess, writers)
+}
+
+// checkDACL refuses an ACL that grants write access to a trustee outside
+// writers.
+//
+// An allowlist, because the trustees that must not have it cannot be listed: an
+// ACE naming an ordinary user account hands that account the same power as one
+// naming Everyone, and only the accounts that may hold it are knowable.
+func checkDACL(path string, dacl *windows.ACL, writeAccess windows.ACCESS_MASK, writers []*windows.SID) error {
+	for i := uint32(0); i < uint32(dacl.AceCount); i++ {
+		var ace *windows.ACCESS_ALLOWED_ACE
+		if err := windows.GetAce(dacl, i, &ace); err != nil {
+			return fmt.Errorf("read ACE %d of %s: %w", i, path, err)
+		}
+		// An inherit-only ACE says what children of this object get, not what
+		// this object grants.
+		if ace.Header.AceFlags&windows.INHERIT_ONLY_ACE != 0 {
+			continue
+		}
+		if ace.Mask&writeAccess == 0 {
+			continue
+		}
+		// Only an allow ACE grants anything; a deny ACE narrows what one gave.
+		if !isAllowACE(ace.Header.AceType) {
+			continue
+		}
+
+		trustee, err := aceTrustee(ace)
+		if err != nil {
+			return fmt.Errorf("read the trustee of ACE %d of %s: %w", i, path, err)
+		}
+		if !containsSID(writers, trustee) {
+			return fmt.Errorf("%s grants write access to %s", path, trustee)
+		}
+	}
+	return nil
+}
+
+// isAllowACE reports whether an ACE type grants rights, rather than denying,
+// auditing or labelling them.
+func isAllowACE(aceType uint8) bool {
+	switch aceType {
+	case windows.ACCESS_ALLOWED_ACE_TYPE, accessAllowedCallbackACEType,
+		accessAllowedObjectACEType, accessAllowedCallbackObjectACEType:
+		return true
+	default:
+		return false
+	}
+}
+
+// aceTrustee returns who an allow ACE grants its rights to. An ACE whose trustee
+// cannot be located is an error rather than something to skip past: being unable
+// to read who is being given write access is a refusal.
+func aceTrustee(ace *windows.ACCESS_ALLOWED_ACE) (*windows.SID, error) {
+	switch ace.Header.AceType {
+	case windows.ACCESS_ALLOWED_ACE_TYPE, accessAllowedCallbackACEType:
+		//nolint:gosec // SidStart is the first uint32 of the variable-length SID that follows the ACE header.
+		return (*windows.SID)(unsafe.Pointer(&ace.SidStart)), nil
+	default:
+		return nil, errors.New("an object-type allow ACE does not carry its trustee where we can read it")
+	}
+}
+
+func containsSID(sids []*windows.SID, sid *windows.SID) bool {
+	return slices.ContainsFunc(sids, sid.Equals)
+}
+
+func currentUserSID() (*windows.SID, error) {
+	token := windows.GetCurrentProcessToken()
+	user, err := token.GetTokenUser()
+	if err != nil {
+		return nil, fmt.Errorf("read this process's user: %w", err)
+	}
+	return user.User.Sid, nil
+}

client/internal/elevate/trusted_windows_test.go

@@ -0,0 +1,126 @@
+package elevate
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/sys/windows"
+)
+
+// A file the test user created under their own profile, which is what a per-user
+// install looks like. The whole chain up to the volume root is walked, so this is
+// also what says the walk does not refuse an ordinary Windows installation: the
+// root of every volume grants BUILTIN\Users rights that are not ours to worry
+// about.
+func TestCheckOnlyOwnerWritableAcceptsOwnFile(t *testing.T) {
+	err := checkOnlyOwnerWritable(writeExecutable(t))
+	assert.NoError(t, err, "a file the test user owns, under directories only administrators can write")
+}
+
+// Write access held by an account that cannot answer the UAC prompt means that
+// account decides what runs behind it, whoever the ACE names. The trustees that
+// must not have it cannot be listed, so the check names the ones that may.
+func TestCheckOnlyOwnerWritableRejectsUntrustedWriters(t *testing.T) {
+	tests := []struct {
+		name      string
+		wellKnown windows.WELL_KNOWN_SID_TYPE
+	}{
+		{name: "everyone", wellKnown: windows.WinWorldSid},
+		{name: "authenticated users", wellKnown: windows.WinAuthenticatedUserSid},
+		{name: "builtin users", wellKnown: windows.WinBuiltinUsersSid},
+		// A service account, which no denylist of the obvious groups would name
+		// and which cannot elevate any more than Everyone can.
+		{name: "local service", wellKnown: windows.WinLocalServiceSid},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			path := writeExecutable(t)
+			grantWrite(t, path, tt.wellKnown)
+
+			assert.Error(t, checkOnlyOwnerWritable(path),
+				"write access for %s must be refused", tt.name)
+		})
+	}
+}
+
+// The masks are the policy: on a file any write reaches its contents, while on a
+// directory only deleting or taking over an entry reaches something already
+// there. Adding an entry does not, which is why the walk survives a volume root.
+func TestWriteAccessMasks(t *testing.T) {
+	assert.NotZero(t, fileWriteAccess&windows.FILE_WRITE_DATA, "writing a file's data reaches its contents")
+	assert.NotZero(t, fileWriteAccess&windows.FILE_APPEND_DATA, "appending to a file reaches its contents")
+
+	assert.Zero(t, dirWriteAccess&windows.FILE_WRITE_DATA, "adding a file to a directory replaces nothing")
+	assert.Zero(t, dirWriteAccess&windows.FILE_APPEND_DATA, "adding a subdirectory replaces nothing")
+	assert.NotZero(t, dirWriteAccess&fileDeleteChild, "deleting an entry replaces it")
+	assert.NotZero(t, dirWriteAccess&windows.DELETE, "deleting the directory takes its entries with it")
+}
+
+func TestIsAllowACE(t *testing.T) {
+	tests := []struct {
+		name    string
+		aceType uint8
+		want    bool
+	}{
+		{name: "allowed", aceType: windows.ACCESS_ALLOWED_ACE_TYPE, want: true},
+		{name: "allowed callback", aceType: accessAllowedCallbackACEType, want: true},
+		{name: "allowed object", aceType: accessAllowedObjectACEType, want: true},
+		{name: "allowed callback object", aceType: accessAllowedCallbackObjectACEType, want: true},
+		{name: "denied", aceType: windows.ACCESS_DENIED_ACE_TYPE},
+		// SYSTEM_AUDIT_ACE_TYPE, which x/sys does not define: an ACE that records
+		// access rather than granting it.
+		{name: "audit", aceType: 0x2},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, isAllowACE(tt.aceType), "ACE type %#x", tt.aceType)
+		})
+	}
+}
+
+// writeExecutable creates a plain file under the test's own directory, the shape
+// trustedSelf checks.
+func writeExecutable(t *testing.T) string {
+	t.Helper()
+	path := filepath.Join(t.TempDir(), "netbird-ui.exe")
+	require.NoError(t, os.WriteFile(path, []byte("MZ"), 0o755), "write the executable")
+	return path
+}
+
+// grantWrite replaces the file's DACL with one that grants a well-known trustee
+// everything, keeping the test user's own access so the file stays deletable.
+func grantWrite(t *testing.T, path string, wellKnown windows.WELL_KNOWN_SID_TYPE) {
+	t.Helper()
+
+	trustee, err := windows.CreateWellKnownSid(wellKnown)
+	require.NoError(t, err, "build the trustee SID")
+	self, err := currentUserSID()
+	require.NoError(t, err, "read the test user's SID")
+
+	acl, err := windows.ACLFromEntries([]windows.EXPLICIT_ACCESS{
+		fullControl(self, windows.TRUSTEE_IS_USER),
+		fullControl(trustee, windows.TRUSTEE_IS_WELL_KNOWN_GROUP),
+	}, nil)
+	require.NoError(t, err, "build the ACL")
+
+	require.NoError(t, windows.SetNamedSecurityInfo(path, windows.SE_FILE_OBJECT,
+		windows.DACL_SECURITY_INFORMATION|windows.PROTECTED_DACL_SECURITY_INFORMATION,
+		nil, nil, acl, nil), "set the DACL")
+}
+
+func fullControl(sid *windows.SID, trusteeType uint32) windows.EXPLICIT_ACCESS {
+	return windows.EXPLICIT_ACCESS{
+		AccessPermissions: windows.GENERIC_ALL,
+		AccessMode:        windows.GRANT_ACCESS,
+		Trustee: windows.TRUSTEE{
+			TrusteeForm:  windows.TRUSTEE_IS_SID,
+			TrusteeType:  windows.TRUSTEE_TYPE(trusteeType),
+			TrusteeValue: windows.TrusteeValueFromSID(sid),
+		},
+	}
+}

client/internal/engine.go

@@ -23,6 +23,7 @@ import (
 	"golang.zx2c4.com/wireguard/tun/netstack"
 	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
 
+	"github.com/netbirdio/netbird/client/anonymize"
 	nberrors "github.com/netbirdio/netbird/client/errors"
 	"github.com/netbirdio/netbird/client/firewall"
 	"github.com/netbirdio/netbird/client/firewall/firewalld"
@@ -59,6 +60,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/syncstore"
 	"github.com/netbirdio/netbird/client/internal/updater"
 	"github.com/netbirdio/netbird/client/jobexec"
+	"github.com/netbirdio/netbird/client/netevents"
 	cProto "github.com/netbirdio/netbird/client/proto"
 	"github.com/netbirdio/netbird/client/system"
 	nbdns "github.com/netbirdio/netbird/dns"
@@ -93,6 +95,13 @@ const (
 	// exec, os.Stat); without this bound a single stuck call freezes handleSync, and
 	// thus syncMsgMux, for as long as the call hangs (observed multi-minute freezes).
 	systemInfoTimeout = 15 * time.Second
+
+	// dnsForwarderStopTimeout bounds how long stopping the DNS forwarder waits
+	// for the queries still in flight. One waiting on an unresponsive upstream
+	// would otherwise hold the stop for the whole upstream timeout, and the
+	// stop runs with syncMsgMux held. The sockets are closed either way, so
+	// giving up costs a query that was already failing.
+	dnsForwarderStopTimeout = 2 * time.Second
 )
 
 var ErrResetConnection = fmt.Errorf("reset connection")
@@ -136,6 +145,7 @@ type EngineConfig struct {
 	RosenpassPermissive bool
 
 	ServerSSHAllowed              bool
+	RemoteJobsAllowed             bool
 	EnableSSHRoot                 *bool
 	EnableSSHSFTP                 *bool
 	EnableSSHLocalPortForwarding  *bool
@@ -181,6 +191,9 @@ type EngineServices struct {
 	UpdateManager  *updater.Manager
 	ClientMetrics  *metrics.ClientMetrics
 	MetricsCtx     context.Context
+	// NetMgr gates the reconnection loops on OS-reported network
+	// availability; nil disables gating.
+	NetMgr *netevents.Manager
 }
 
 // Engine is a mechanism responsible for reacting on Signal and Management stream events and managing connections to the remote peers.
@@ -204,6 +217,10 @@ type Engine struct {
 	config    *EngineConfig
 	mobileDep MobileDependency
 
+	// netMgr gates the peer reconnection guards on OS-reported network
+	// availability; nil disables gating.
+	netMgr *netevents.Manager
+
 	// STUNs is a list of STUN servers used by ICE
 	STUNs []*stun.URI
 	// TURNs is a list of STUN servers used by ICE
@@ -249,6 +266,8 @@ type Engine struct {
 	// checks are the client-applied posture checks that need to be evaluated on the client
 	checks []*mgmProto.Checks
 
+	infoSource system.InfoSource
+
 	relayManager       *relayClient.Manager
 	stateManager       *statemanager.Manager
 	portForwardManager *portforward.Manager
@@ -312,6 +331,10 @@ type localIpUpdater interface {
 	UpdateLocalIPs() error
 }
 
+// overlayRebind rebuilds one subsystem's sockets on the current interface. The
+// error it returns names its own subsystem, since the caller can only log it.
+type overlayRebind func() error
+
 // NewEngine creates a new Connection Engine with probes attached
 func NewEngine(
 	clientCtx context.Context,
@@ -337,6 +360,7 @@ func NewEngine(
 		syncMsgMux:         &sync.Mutex{},
 		config:             config,
 		mobileDep:          mobileDep,
+		netMgr:             services.NetMgr,
 		STUNs:              []*stun.URI{},
 		TURNs:              []*stun.URI{},
 		networkSerial:      0,
@@ -735,6 +759,11 @@ func (e *Engine) initFirewall() error {
 		return fmt.Errorf("set firewall: %w", err)
 	}
 
+	// TODO: the firewall backends dedup filter rules by content, so a
+	// management route ACL with identical content would collapse onto the
+	// untracked drop rules installed here, and a later management delete
+	// could remove them. Needs backend refcounting or per-consumer key
+	// namespacing.
 	if e.config.BlockLANAccess {
 		e.blockLanAccess()
 	}
@@ -747,14 +776,14 @@ func (e *Engine) initFirewall() error {
 	port := firewallManager.Port{Values: []uint16{uint16(rosenpassPort)}}
 
 	// IPv4-only: rosenpass peers connect via AllowedIps[0] which is always v4.
-	if _, err := e.firewall.AddPeerFiltering(
+	if _, err := e.firewall.AddFilterRule(
 		nil,
-		net.IP{0, 0, 0, 0},
+		[]netip.Prefix{netip.PrefixFrom(netip.IPv4Unspecified(), 0)},
+		firewallManager.Network{},
 		firewallManager.ProtocolUDP,
 		nil,
 		&port,
 		firewallManager.ActionAccept,
-		"",
 	); err != nil {
 		log.Errorf("failed to allow rosenpass interface traffic: %v", err)
 		return nil
@@ -804,7 +833,7 @@ func (e *Engine) blockLanAccess() {
 		if network.Addr().Is6() {
 			source = v6
 		}
-		if _, err := e.firewall.AddRouteFiltering(
+		if _, err := e.firewall.AddFilterRule(
 			nil,
 			[]netip.Prefix{source},
 			firewallManager.Network{Prefix: network},
@@ -863,8 +892,7 @@ func (e *Engine) modifyPeers(peersUpdate []*mgmProto.RemotePeerConfig) error {
 	}
 	// third, add the peer connections again
 	for _, p := range modified {
-		err := e.addNewPeer(p)
-		if err != nil {
+		if err := e.addNewPeer(p); err != nil {
 			return err
 		}
 	}
@@ -1216,9 +1244,7 @@ func (e *Engine) updateChecksIfNew(checks []*mgmProto.Checks) error {
 	if isChecksEqual(e.checks, checks) {
 		return nil
 	}
-	e.checks = checks
-
-	info, ok := system.GetInfoWithChecksTimeout(e.ctx, systemInfoTimeout, checks, e.overlayAddresses()...)
+	info, ok := e.infoSource.Refresh(e.ctx, systemInfoTimeout, checks, e.overlayAddresses()...)
 	if !ok {
 		// Gathering timed out; skip the meta sync this cycle rather than blocking the
 		// sync loop (and syncMsgMux) on a stuck system call. A later sync will retry.
@@ -1229,6 +1255,7 @@ func (e *Engine) updateChecksIfNew(checks []*mgmProto.Checks) error {
 	if err := e.mgmClient.SyncMeta(info); err != nil {
 		return fmt.Errorf("could not sync meta: error %s", err)
 	}
+	e.checks = checks
 	return nil
 }
 
@@ -1251,9 +1278,32 @@ func (e *Engine) applyInfoFlags(info *system.Info) {
 		e.config.EnableSSHLocalPortForwarding,
 		e.config.EnableSSHRemotePortForwarding,
 		e.config.DisableSSHAuth,
+		&e.config.RemoteJobsAllowed,
 	)
 }
 
+func (e *Engine) currentSystemInfo(ctx context.Context) *system.Info {
+	info := e.infoSource.Current(ctx, e.overlayAddresses()...)
+	e.applyInfoFlags(info)
+	return info
+}
+
+// syncInfoFunc returns the info callback for the management sync stream. The
+// first connect sends the info refreshed right before it instead of gathering
+// again; every reconnect gathers a fresh one. The stream retry loop calls the
+// callback sequentially, so the handoff needs no synchronization.
+func (e *Engine) syncInfoFunc(refreshed *system.Info) func(ctx context.Context) *system.Info {
+	return func(ctx context.Context) *system.Info {
+		if refreshed == nil {
+			return e.currentSystemInfo(ctx)
+		}
+		info := refreshed
+		refreshed = nil
+		e.applyInfoFlags(info)
+		return info
+	}
+}
+
 // overlayAddresses returns our own WireGuard overlay address (v4 and v6) so it
 // can be excluded from the reported network addresses; the interface coming and
 // going otherwise churns the peer meta on the management server.
@@ -1336,6 +1386,13 @@ func (e *Engine) receiveJobEvents() {
 				ID:     msg.ID,
 				Status: mgmProto.JobStatus_failed,
 			}
+			// Remote jobs are an explicit opt-in. When not enabled on this
+			// peer, every job is refused before any work is done.
+			if !e.config.RemoteJobsAllowed {
+				log.Warnf("refusing remote job: remote jobs are not enabled on this peer (enable with --allow-remote-jobs)")
+				resp.Reason = []byte("remote jobs are not enabled on this peer")
+				return &resp
+			}
 			switch params := msg.WorkloadParameters.(type) {
 			case *mgmProto.JobRequest_Bundle:
 				bundleResult, err := e.handleBundle(params.Bundle)
@@ -1365,7 +1422,25 @@ func (e *Engine) receiveJobEvents() {
 }
 
 func (e *Engine) handleBundle(params *mgmProto.BundleParameters) (*mgmProto.JobResponse_Bundle, error) {
-	log.Infof("handle remote debug bundle request: %s", params.String())
+	// The upload URL can carry a host, credentials, or query tokens, so it is
+	// kept out of the info-level line; the full parameters stay available at
+	// debug level for troubleshooting.
+	log.Infof("handle remote debug bundle request: anonymize=%v anonymize_level=%q log_file_count=%d bundle_for=%v bundle_for_time=%d",
+		params.GetAnonymize(), params.GetAnonymizeLevel(), params.GetLogFileCount(), params.GetBundleFor(), params.GetBundleForTime())
+	log.Debugf("remote debug bundle request parameters: %s", params.String())
+
+	// Resolve the upload destination: an MDM override, when set, takes
+	// precedence over the management-supplied URL. Both are validated the same
+	// way; an empty result falls back to the default upload server downstream.
+	uploadURL := params.GetUploadUrl()
+	if override := e.config.ProfileConfig.DebugBundleUploadURL; override != "" {
+		log.Infof("using MDM debug bundle upload URL override instead of the management-supplied value")
+		uploadURL = override
+	}
+	if err := validateBundleUploadURL(uploadURL); err != nil {
+		return nil, err
+	}
+
 	syncResponse, err := e.GetLatestSyncResponse()
 	if err != nil {
 		log.Warnf("get latest sync response: %v", err)
@@ -1386,13 +1461,14 @@ func (e *Engine) handleBundle(params *mgmProto.BundleParameters) (*mgmProto.JobR
 
 	bundleJobParams := debug.BundleConfig{
 		Anonymize:         params.Anonymize,
+		AnonymizeLevel:    anonymize.ParseLevel(params.AnonymizeLevel),
 		IncludeSystemInfo: true,
 		LogFileCount:      uint32(params.LogFileCount),
 	}
 
 	waitFor := time.Duration(params.BundleForTime) * time.Minute
 
-	uploadKey, err := e.jobExecutor.BundleJob(e.ctx, bundleDeps, bundleJobParams, waitFor, e.config.ProfileConfig.ManagementURL.String())
+	uploadKey, err := e.jobExecutor.BundleJob(e.ctx, bundleDeps, bundleJobParams, waitFor, e.config.ProfileConfig.ManagementURL.String(), uploadURL)
 	if err != nil {
 		return nil, err
 	}
@@ -1405,21 +1481,27 @@ func (e *Engine) handleBundle(params *mgmProto.BundleParameters) (*mgmProto.JobR
 	return response, nil
 }
 
+// validateBundleUploadURL sanity-checks a management-supplied upload URL for a
+// remote debug bundle job. It delegates to profilemanager.ValidateBundleUploadURL
+// so the executor and the MDM policy override share one definition of the rule
+// (empty accepted; otherwise a well-formed https URL with a host) and cannot
+// drift. The host is deliberately left unconstrained pending a decision on
+// management-directed uploads.
+func validateBundleUploadURL(raw string) error {
+	return profilemanager.ValidateBundleUploadURL(raw)
+}
+
 // receiveManagementEvents connects to the Management Service event stream to receive updates from the management service
 // E.g. when a new peer has been registered and we are allowed to connect to it.
 func (e *Engine) receiveManagementEvents() {
 	e.shutdownWg.Add(1)
 	go func() {
 		defer e.shutdownWg.Done()
-		info, ok := system.GetInfoWithChecksTimeout(e.ctx, systemInfoTimeout, e.checks, e.overlayAddresses()...)
+		info, ok := e.infoSource.Refresh(e.ctx, systemInfoTimeout, e.checks, e.overlayAddresses()...)
 		if !ok {
-			// Gathering timed out; connect the stream with base info so management
-			// connectivity still comes up rather than blocking here.
-			info = system.GetInfo(e.ctx)
+			log.Warnf("posture checks not refreshed before the sync connect, sending the previous results")
 		}
-		e.applyInfoFlags(info)
-
-		err := e.mgmClient.Sync(e.ctx, info, e.handleSync)
+		err := e.mgmClient.Sync(e.ctx, e.syncInfoFunc(info), e.handleSync)
 		if err != nil {
 			// happens if management is unavailable for a long time.
 			// We want to cancel the operation of the whole client
@@ -1485,8 +1567,12 @@ func (e *Engine) updateNetworkMap(networkMap *mgmProto.NetworkMap) error {
 		return nil
 	}
 
-	if err := e.connMgr.UpdatedRemoteFeatureFlag(e.ctx, networkMap.GetPeerConfig().GetLazyConnectionEnabled()); err != nil {
-		log.Errorf("failed to update lazy connection feature flag: %v", err)
+	// Only update the flag when the sync carries a peer config; a nil peer config
+	// (e.g. a partial update) must not reset the cached flag to false.
+	if peerConfig := networkMap.GetPeerConfig(); peerConfig != nil {
+		if err := e.connMgr.UpdatedRemoteFeatureFlag(e.ctx, peerConfig.GetLazyConnectionEnabled()); err != nil {
+			log.Errorf("failed to update lazy connection feature flag: %v", err)
+		}
 	}
 
 	if e.firewall != nil {
@@ -1552,8 +1638,7 @@ func (e *Engine) updateNetworkMap(networkMap *mgmProto.NetworkMap) error {
 
 	// Ingress forward rules
 	done = e.phase("forward_rules")
-	forwardingRules, err := e.updateForwardRules(networkMap.GetForwardingRules())
-	if err != nil {
+	if _, err := e.updateForwardRules(networkMap.GetForwardingRules()); err != nil {
 		log.Errorf("failed to update forward rules, err: %v", err)
 	}
 	done()
@@ -1571,8 +1656,7 @@ func (e *Engine) updateNetworkMap(networkMap *mgmProto.NetworkMap) error {
 
 	// must set the exclude list after the peers are added. Without it the manager can not figure out the peers parameters from the store
 	done = e.phase("lazy_exclude")
-	excludedLazyPeers := e.toExcludedLazyPeers(forwardingRules, remotePeers)
-	e.connMgr.SetExcludeList(e.ctx, excludedLazyPeers)
+	e.connMgr.SetExcludeList(e.ctx, e.toExcludedLazyPeers(remotePeers))
 	done()
 
 	e.networkSerial = serial
@@ -1816,15 +1900,15 @@ func addrToString(addr netip.Addr) string {
 // addNewPeers adds peers that were not know before but arrived from the Management service with the update
 func (e *Engine) addNewPeers(peersUpdate []*mgmProto.RemotePeerConfig) error {
 	for _, p := range peersUpdate {
-		err := e.addNewPeer(p)
-		if err != nil {
+		if err := e.addNewPeer(p); err != nil {
 			return err
 		}
 	}
 	return nil
 }
 
-// addNewPeer add peer if connection doesn't exist
+// addNewPeer add peer if connection doesn't exist. A peer that is not lazy by
+// policy gets an always-active connection instead.
 func (e *Engine) addNewPeer(peerConfig *mgmProto.RemotePeerConfig) error {
 	peerKey := peerConfig.GetWgPubKey()
 	peerIPs := make([]netip.Prefix, 0, len(peerConfig.GetAllowedIps()))
@@ -1859,7 +1943,8 @@ func (e *Engine) addNewPeer(peerConfig *mgmProto.RemotePeerConfig) error {
 		log.Warnf("error adding peer %s to status recorder, got error: %v", peerKey, err)
 	}
 
-	if exists := e.connMgr.AddPeerConn(e.ctx, peerKey, conn); exists {
+	permanent := !e.connMgr.PeerLazyDefault(peerConfig.GetLazyState())
+	if exists := e.connMgr.AddPeerConn(e.ctx, peerKey, conn, permanent); exists {
 		conn.Close(false)
 		return fmt.Errorf("peer already exists: %s", peerKey)
 	}
@@ -1893,6 +1978,7 @@ func (e *Engine) createPeerConn(pubKey string, allowedIPs []netip.Prefix, agentV
 			PermissiveMode: e.config.RosenpassPermissive,
 		},
 		ICEConfig: e.createICEConfig(),
+		NetMgr:    e.netMgr,
 	}
 
 	serviceDependencies := peer.ServiceDependencies{
@@ -2447,7 +2533,72 @@ func (e *Engine) RenewTun(fd int) error {
 		return fmt.Errorf("wireguard interface not initialized")
 	}
 
-	return wgInterface.RenewTun(fd)
+	if err := wgInterface.RenewTun(fd); err != nil {
+		return err
+	}
+
+	e.rebindOverlayListeners()
+	return nil
+}
+
+// rebindOverlayListeners gives the servers that listen on an overlay address
+// sockets on the interface as it is now.
+//
+// A socket belongs to the interface generation it was created on. Renewing the
+// TUN builds a new interface and moves the overlay addresses to it, which
+// leaves the old sockets in LISTEN with the uspfilter still logging packets
+// arriving for them, while every accept fails with EINVAL for the life of the
+// socket: from the outside the server looks alive and answers nothing. On
+// Android this happens during a normal startup, where the first TUN is
+// established before the routes are known and replaced once they arrive.
+//
+// Rebinding costs whatever those sockets were carrying, which the renewal has
+// already broken. Errors are logged rather than returned: the renewal itself
+// succeeded, and failing it would hand the caller a working interface and an
+// error.
+func (e *Engine) rebindOverlayListeners() {
+	e.syncMsgMux.Lock()
+	defer e.syncMsgMux.Unlock()
+
+	for _, rebind := range e.overlayRebinds() {
+		if err := rebind(); err != nil {
+			log.Errorf("after TUN renewal: %v", err)
+		}
+	}
+}
+
+// overlayRebinds is every subsystem of this engine that holds sockets bound to
+// an overlay address, and how to rebuild each one's.
+//
+// A subsystem that starts listening on an overlay address belongs in this list.
+// Leaving it out costs nothing that review would notice and produces a listener
+// that stays in LISTEN, is logged as receiving packets, and refuses every
+// connection for the life of the process.
+func (e *Engine) overlayRebinds() []overlayRebind {
+	return []overlayRebind{
+		e.restartSSHListeners,
+		e.restartDNSForwarder,
+	}
+}
+
+// restartDNSForwarder rebuilds the DNS forwarder serving the same domains.
+// No-op when it is not running. See Engine.rebindOverlayListeners.
+func (e *Engine) restartDNSForwarder() error {
+	if e.dnsForwardMgr == nil {
+		return nil
+	}
+	// Read from the forwarder before it goes away, so the replacement serves
+	// the domains in force now rather than a copy kept somewhere else.
+	entries := e.dnsForwardMgr.Domains()
+	e.stopDNSForwarder()
+	// Both halves log their own failures, so the only thing left to report is
+	// the outcome: a start that failed left the manager nil, and the forwarder
+	// is now down rather than merely rebound.
+	e.startDNSForwarder(entries)
+	if e.dnsForwardMgr == nil {
+		return errors.New("rebind DNS forwarder: it did not come back up")
+	}
+	return nil
 }
 
 // updateDNSForwarder start or stop the DNS forwarder based on the domains and the feature flag
@@ -2493,7 +2644,14 @@ func (e *Engine) stopDNSForwarder() {
 		return
 	}
 
-	if err := e.dnsForwardMgr.Stop(context.Background()); err != nil {
+	// Bounded because the shutdown waits for queries still in flight, and one
+	// waiting on an unresponsive upstream holds it for as long as that lookup
+	// is allowed to take. This runs with syncMsgMux held, so that wait is one
+	// the whole engine spends.
+	ctx, cancel := context.WithTimeout(context.Background(), dnsForwarderStopTimeout)
+	defer cancel()
+
+	if err := e.dnsForwardMgr.Stop(ctx); err != nil {
 		log.Errorf("failed to stop DNS forward: %v", err)
 	}
 
@@ -2561,7 +2719,7 @@ func (e *Engine) SetCapture(pc device.PacketCapture) error {
 	}
 
 	afc := capture.NewAFPacketCapture(intf.Name(), sess)
-	if err := afc.Start(); err != nil {
+	if err := afc.Start(); err != nil { //nolint:staticcheck // always errors on non-Linux builds
 		return fmt.Errorf("start AF_PACKET capture on %s: %w", intf.Name(), err)
 	}
 	e.afpacketCapture = afc
@@ -2608,7 +2766,7 @@ func (e *Engine) updateForwardRules(rules []*mgmProto.ForwardingRule) ([]firewal
 	var merr *multierror.Error
 	forwardingRules := make([]firewallManager.ForwardRule, 0, len(rules))
 	for _, rule := range rules {
-		proto, err := convertToFirewallProtocol(rule.GetProtocol())
+		proto, err := acl.ConvertToFirewallProtocol(rule.GetProtocol())
 		if err != nil {
 			merr = multierror.Append(merr, fmt.Errorf("failed to convert protocol '%s': %w", rule.GetProtocol(), err))
 			continue
@@ -2650,46 +2808,19 @@ func (e *Engine) updateForwardRules(rules []*mgmProto.ForwardingRule) ([]firewal
 	return forwardingRules, nberrors.FormatErrorOrNil(merr)
 }
 
-func (e *Engine) toExcludedLazyPeers(rules []firewallManager.ForwardRule, peers []*mgmProto.RemotePeerConfig) map[string]bool {
+// toExcludedLazyPeers returns the peers that must have an always-active
+// connection: those that are not lazy by policy (the per-peer lazy state or the
+// account flag, subject to the local override).
+func (e *Engine) toExcludedLazyPeers(peers []*mgmProto.RemotePeerConfig) map[string]bool {
 	excludedPeers := make(map[string]bool)
-
-	// Ingress forward targets: inbound forwarded traffic is initiated remotely and
-	// cannot wake a lazy connection, so the peer routing the target must stay
-	// permanently connected. AllowedIPs are already parsed on the peer conn, so
-	// reuse those typed prefixes instead of re-parsing the network map strings.
-	for _, r := range rules {
-		for _, p := range peers {
-			if e.peerRoutesAddr(p, r.TranslatedAddress) {
-				log.Infof("exclude forwarder peer from lazy connection: %s", p.GetWgPubKey())
-				excludedPeers[p.GetWgPubKey()] = true
-			}
+	for _, p := range peers {
+		if !e.connMgr.PeerLazyDefault(p.GetLazyState()) {
+			excludedPeers[p.GetWgPubKey()] = true
 		}
 	}
-
 	return excludedPeers
 }
 
-// peerRoutesAddr reports whether the peer is a router for addr, matched against
-// the peer's already-parsed AllowedIPs from the store (the same typed value the
-// lazy manager consumes) rather than re-parsing the network map strings.
-func (e *Engine) peerRoutesAddr(p *mgmProto.RemotePeerConfig, addr netip.Addr) bool {
-	prefixes, ok := e.peerStore.AllowedIPs(p.GetWgPubKey())
-	if !ok {
-		return false
-	}
-	return prefixesContain(prefixes, addr)
-}
-
-// prefixesContain reports whether addr falls within any of the prefixes.
-func prefixesContain(prefixes []netip.Prefix, addr netip.Addr) bool {
-	for _, prefix := range prefixes {
-		if prefix.Contains(addr) {
-			return true
-		}
-	}
-	return false
-}
-
 // isChecksEqual checks if two slices of checks are equal.
 func isChecksEqual(checks1, checks2 []*mgmProto.Checks) bool {
 	normalize := func(checks []*mgmProto.Checks) []string {

client/internal/engine_bundle_test.go

@@ -0,0 +1,36 @@
+package internal
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestValidateBundleUploadURL covers the sanity check applied to a
+// management-supplied upload URL before a remote debug bundle is generated.
+func TestValidateBundleUploadURL(t *testing.T) {
+	for _, tc := range []struct {
+		name    string
+		raw     string
+		wantErr bool
+	}{
+		{name: "empty falls back to default", raw: ""},
+		{name: "https with host", raw: "https://upload.debug.netbird.io/upload"},
+		{name: "https self-hosted host", raw: "https://upload.example.com"},
+		{name: "plaintext rejected", raw: "http://upload.example.com", wantErr: true},
+		{name: "missing host rejected", raw: "https:///upload", wantErr: true},
+		{name: "port-only authority rejected", raw: "https://:443", wantErr: true},
+		{name: "non-url scheme rejected", raw: "ftp://upload.example.com", wantErr: true},
+		{name: "garbage rejected", raw: "://not a url", wantErr: true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			err := validateBundleUploadURL(tc.raw)
+			if tc.wantErr {
+				require.Error(t, err, "an invalid upload URL must be rejected")
+				return
+			}
+			assert.NoError(t, err, "a valid or empty upload URL must be accepted")
+		})
+	}
+}

client/internal/engine_lazy_exclude_test.go

@@ -1,87 +0,0 @@
-package internal
-
-import (
-	"net/netip"
-	"testing"
-
-	"github.com/stretchr/testify/require"
-
-	firewallManager "github.com/netbirdio/netbird/client/firewall/manager"
-	"github.com/netbirdio/netbird/client/internal/peer"
-	"github.com/netbirdio/netbird/client/internal/peerstore"
-	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
-)
-
-func TestPrefixesContain(t *testing.T) {
-	tests := []struct {
-		name     string
-		prefixes []string
-		addr     string
-		want     bool
-	}{
-		{name: "own overlay /32 matches", prefixes: []string{"100.110.8.145/32"}, addr: "100.110.8.145", want: true},
-		{name: "addr inside routed subnet", prefixes: []string{"10.121.0.0/16"}, addr: "10.121.208.4", want: true},
-		{name: "addr outside subnet", prefixes: []string{"10.121.0.0/16"}, addr: "10.122.0.1", want: false},
-		{name: "different /32", prefixes: []string{"100.110.8.145/32"}, addr: "100.110.8.146", want: false},
-		{name: "ipv6 /128 matches", prefixes: []string{"fd00::1/128"}, addr: "fd00::1", want: true},
-		{name: "no prefixes", prefixes: nil, addr: "10.121.208.4", want: false},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			prefixes := make([]netip.Prefix, 0, len(tt.prefixes))
-			for _, p := range tt.prefixes {
-				prefixes = append(prefixes, netip.MustParsePrefix(p))
-			}
-			require.Equal(t, tt.want, prefixesContain(prefixes, netip.MustParseAddr(tt.addr)))
-		})
-	}
-}
-
-// TestToExcludedLazyPeers_ForwardTarget guards a regression: the forward-target
-// peer (the peer routing a ForwardRule.TranslatedAddress) must be excluded from
-// lazy connections, matched via the peer's already-parsed AllowedIPs.
-func TestToExcludedLazyPeers_ForwardTarget(t *testing.T) {
-	const targetPeerKey = "cccccccccccccccccccccccccccccccccccccccccc0="
-	const otherPeerKey = "dddddddddddddddddddddddddddddddddddddddddd0="
-
-	store := peerstore.NewConnStore()
-	store.AddPeerConn(targetPeerKey, newTestConn(t, targetPeerKey, "100.110.8.145/32"))
-	store.AddPeerConn(otherPeerKey, newTestConn(t, otherPeerKey, "100.110.9.10/32"))
-
-	e := &Engine{peerStore: store}
-
-	peers := []*mgmProto.RemotePeerConfig{
-		{WgPubKey: targetPeerKey, AllowedIps: []string{"100.110.8.145/32"}},
-		{WgPubKey: otherPeerKey, AllowedIps: []string{"100.110.9.10/32"}},
-	}
-	rules := []firewallManager.ForwardRule{
-		{TranslatedAddress: netip.MustParseAddr("100.110.8.145")},
-	}
-
-	excluded := e.toExcludedLazyPeers(rules, peers)
-
-	require.True(t, excluded[targetPeerKey], "forward-target peer must be excluded from lazy connections")
-	require.False(t, excluded[otherPeerKey], "non-target peer must not be excluded")
-	require.Len(t, excluded, 1)
-}
-
-func TestToExcludedLazyPeers_NoRules(t *testing.T) {
-	e := &Engine{peerStore: peerstore.NewConnStore()}
-
-	peers := []*mgmProto.RemotePeerConfig{
-		{WgPubKey: "peer-a", AllowedIps: []string{"100.110.8.145/32"}},
-	}
-
-	require.Empty(t, e.toExcludedLazyPeers(nil, peers))
-}
-
-func newTestConn(t *testing.T, key, allowedIP string) *peer.Conn {
-	t.Helper()
-	conn, err := peer.NewConn(peer.ConnConfig{
-		Key:      key,
-		WgConfig: peer.WgConfig{AllowedIps: []netip.Prefix{netip.MustParsePrefix(allowedIP)}},
-	}, peer.ServiceDependencies{})
-	require.NoError(t, err)
-	return conn
-}

client/internal/engine_privileged_test.go

@@ -12,7 +12,7 @@ import (
 	"testing"
 	"time"
 
-	"github.com/golang/mock/gomock"
+	"go.uber.org/mock/gomock"
 	"github.com/google/uuid"
 	log "github.com/sirupsen/logrus"
 	"github.com/stretchr/testify/assert"
@@ -193,7 +193,7 @@ func TestEngine_Sync(t *testing.T) {
 	// feed updates to Engine via mocked Management client
 	updates := make(chan *mgmtProto.SyncResponse)
 	defer close(updates)
-	syncFunc := func(ctx context.Context, info *system.Info, msgHandler func(msg *mgmtProto.SyncResponse) error) error {
+	syncFunc := func(ctx context.Context, _ func(context.Context) *system.Info, msgHandler func(msg *mgmtProto.SyncResponse) error) error {
 		for msg := range updates {
 			err := msgHandler(msg)
 			if err != nil {
@@ -519,7 +519,7 @@ func startManagement(t *testing.T, dataDir, testFile string) (*grpc.Server, stri
 
 	updateManager := update_channel.NewPeersUpdateManager(metrics)
 	requestBuffer := server.NewAccountRequestBuffer(context.Background(), store)
-	networkMapController := controller.NewController(context.Background(), store, metrics, updateManager, requestBuffer, server.MockIntegratedValidator{}, settingsMockManager, "netbird.selfhosted", port_forwarding.NewControllerMock(), manager.NewEphemeralManager(store, peersManager), config)
+	networkMapController := controller.NewController(context.Background(), store, metrics, updateManager, requestBuffer, server.MockIntegratedValidator{}, settingsMockManager, "netbird.selfhosted", port_forwarding.NewControllerMock(), manager.NewEphemeralManager(store, peersManager), config, nil)
 	accountManager, err := server.BuildManager(context.Background(), config, store, networkMapController, jobManager, nil, "", eventStore, nil, false, ia, metrics, port_forwarding.NewControllerMock(), settingsMockManager, permissionsManager, false, cacheStore)
 	if err != nil {
 		return nil, "", err

client/internal/engine_ssh.go

@@ -24,6 +24,8 @@ type sshServer interface {
 	Stop() error
 	GetStatus() (bool, []sshserver.SessionInfo)
 	UpdateSSHAuth(config *sshauth.Config)
+	JWTConfig() *sshserver.JWTConfig
+	AuthConfig() *sshauth.Config
 }
 
 func (e *Engine) setupSSHPortRedirection() error {
@@ -77,7 +79,7 @@ func (e *Engine) updateSSH(sshConf *mgmProto.SSHConfig) error {
 
 	if e.config.DisableSSHAuth != nil && *e.config.DisableSSHAuth {
 		log.Info("starting SSH server without JWT authentication (authentication disabled by config)")
-		return e.startSSHServer(nil)
+		return e.startSSHServer(nil, nil)
 	}
 
 	if protoJWT := sshConf.GetJwtConfig(); protoJWT != nil {
@@ -95,7 +97,7 @@ func (e *Engine) updateSSH(sshConf *mgmProto.SSHConfig) error {
 			MaxTokenAge:  protoJWT.GetMaxTokenAge(),
 		}
 
-		return e.startSSHServer(jwtConfig)
+		return e.startSSHServer(jwtConfig, nil)
 	}
 
 	return errors.New("SSH server requires valid JWT configuration")
@@ -231,8 +233,33 @@ func (e *Engine) cleanupSSHConfig() {
 	}
 }
 
-// startSSHServer initializes and starts the SSH server with proper configuration.
-func (e *Engine) startSSHServer(jwtConfig *sshserver.JWTConfig) error {
+// restartSSHListeners rebuilds the SSH server so it listens on new sockets, on
+// the same terms it was started with. No-op when it is not running. See
+// Engine.rebindOverlayListeners for why this is needed.
+func (e *Engine) restartSSHListeners() error {
+	if e.sshServer == nil {
+		return nil
+	}
+	// Read from the server before it goes away. A rebuilt one starts with an
+	// empty authorizer, which fails closed, so without carrying the
+	// authorization over every JWT login is refused until the next network map
+	// happens to bring one.
+	jwtConfig, authConfig := e.sshServer.JWTConfig(), e.sshServer.AuthConfig()
+	if err := e.stopSSHServer(); err != nil {
+		return fmt.Errorf("rebind SSH listeners: %w", err)
+	}
+	if err := e.startSSHServer(jwtConfig, authConfig); err != nil {
+		return fmt.Errorf("rebind SSH listeners: %w", err)
+	}
+	return nil
+}
+
+// startSSHServer initializes and starts the SSH server with proper
+// configuration. authConfig is the fine-grained authorization to open with, and
+// is applied before the server accepts anything: a server that starts listening
+// with an empty authorizer refuses the logins that arrive in the meantime.
+// Nil leaves it as management has not sent one yet.
+func (e *Engine) startSSHServer(jwtConfig *sshserver.JWTConfig, authConfig *sshauth.Config) error {
 	if e.wgInterface == nil {
 		return errors.New("wg interface not initialized")
 	}
@@ -240,6 +267,7 @@ func (e *Engine) startSSHServer(jwtConfig *sshserver.JWTConfig) error {
 	serverConfig := &sshserver.Config{
 		HostKeyPEM: e.config.SSHKey,
 		JWT:        jwtConfig,
+		Auth:       authConfig,
 	}
 	server := sshserver.New(serverConfig)
 

client/internal/engine_test.go

@@ -2,6 +2,7 @@ package internal
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"net"
 	"net/netip"
@@ -31,6 +32,7 @@ import (
 	icemaker "github.com/netbirdio/netbird/client/internal/peer/ice"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/internal/routemanager"
+	"github.com/netbirdio/netbird/client/system"
 	nbdns "github.com/netbirdio/netbird/dns"
 	"github.com/netbirdio/netbird/monotime"
 	"github.com/netbirdio/netbird/route"
@@ -253,6 +255,118 @@ func TestEngine_SSHServerConsistency(t *testing.T) {
 	})
 }
 
+func TestEngine_FirstSyncInfoCarriesLoginChecks(t *testing.T) {
+	key, err := wgtypes.GeneratePrivateKey()
+	require.NoError(t, err)
+
+	exe, err := os.Executable()
+	require.NoError(t, err)
+
+	ctx, cancel := context.WithCancel(CtxInitState(context.Background()))
+	defer cancel()
+
+	infos := make(chan *system.Info, 1)
+	mgmClient := &mgmt.MockClient{
+		SyncFunc: func(ctx context.Context, getInfo func(context.Context) *system.Info, _ func(*mgmtProto.SyncResponse) error) error {
+			infos <- getInfo(ctx)
+			return nil
+		},
+	}
+
+	relayMgr := relayClient.NewManager(ctx, nil, key.PublicKey().String(), iface.DefaultMTU)
+	engine := NewEngine(ctx, cancel, &EngineConfig{
+		WgIfaceName:  "utun104",
+		WgAddr:       wgaddr.MustParseWGAddress("100.64.0.1/24"),
+		WgPrivateKey: key,
+		WgPort:       33100,
+		MTU:          iface.DefaultMTU,
+	}, EngineServices{
+		SignalClient:   &signal.MockClient{},
+		MgmClient:      mgmClient,
+		RelayManager:   relayMgr,
+		StatusRecorder: peer.NewRecorder("https://mgm"),
+		Checks:         []*mgmtProto.Checks{{Files: []string{exe}}},
+	}, MobileDependency{})
+
+	engine.receiveManagementEvents()
+
+	select {
+	case info := <-infos:
+		require.Len(t, info.Files, 1)
+		assert.Equal(t, exe, info.Files[0].Path)
+		assert.True(t, info.Files[0].Exist)
+	case <-time.After(20 * time.Second):
+		t.Fatal("timeout waiting for the first sync info")
+	}
+	engine.shutdownWg.Wait()
+}
+
+func TestEngine_SyncInfoFuncReusesRefreshedInfoOnce(t *testing.T) {
+	engine := &Engine{config: &EngineConfig{}}
+
+	refreshed := &system.Info{Hostname: "from-refresh"}
+	getInfo := engine.syncInfoFunc(refreshed)
+
+	first := getInfo(context.Background())
+	assert.Same(t, refreshed, first, "the first connect should send the refreshed info instead of gathering again")
+
+	second := getInfo(context.Background())
+	assert.NotSame(t, refreshed, second, "the reconnect should gather a fresh info")
+	assert.NotEqual(t, "from-refresh", second.Hostname, "the fresh info should not carry the refreshed hostname")
+}
+
+func TestEngine_SyncInfoFuncGathersWhenRefreshFailed(t *testing.T) {
+	engine := &Engine{config: &EngineConfig{}}
+
+	info := engine.syncInfoFunc(nil)(context.Background())
+	require.NotNil(t, info, "a failed refresh should fall back to gathering the info")
+	assert.NotEmpty(t, info.Hostname, "the gathered info should carry the hostname")
+}
+
+func TestEngine_UpdateChecksIfNewRetriesAfterFailedSyncMeta(t *testing.T) {
+	key, err := wgtypes.GeneratePrivateKey()
+	require.NoError(t, err)
+
+	exe, err := os.Executable()
+	require.NoError(t, err)
+
+	ctx, cancel := context.WithCancel(CtxInitState(context.Background()))
+	defer cancel()
+
+	syncMetaCalls := 0
+	mgmClient := &mgmt.MockClient{
+		SyncMetaFunc: func(*system.Info) error {
+			syncMetaCalls++
+			if syncMetaCalls == 1 {
+				return errors.New("management unavailable")
+			}
+			return nil
+		},
+	}
+
+	relayMgr := relayClient.NewManager(ctx, nil, key.PublicKey().String(), iface.DefaultMTU)
+	engine := NewEngine(ctx, cancel, &EngineConfig{
+		WgIfaceName:  "utun105",
+		WgAddr:       wgaddr.MustParseWGAddress("100.64.0.1/24"),
+		WgPrivateKey: key,
+		WgPort:       33100,
+		MTU:          iface.DefaultMTU,
+	}, EngineServices{
+		SignalClient:   &signal.MockClient{},
+		MgmClient:      mgmClient,
+		RelayManager:   relayMgr,
+		StatusRecorder: peer.NewRecorder("https://mgm"),
+	}, MobileDependency{})
+
+	checks := []*mgmtProto.Checks{{Files: []string{exe}}}
+
+	require.Error(t, engine.updateChecksIfNew(checks))
+	require.NoError(t, engine.updateChecksIfNew(checks))
+	require.NoError(t, engine.updateChecksIfNew(checks))
+
+	assert.Equal(t, 2, syncMetaCalls)
+}
+
 func TestEngine_UpdateNetworkMap(t *testing.T) {
 	// test setup
 	key, err := wgtypes.GeneratePrivateKey()
@@ -279,7 +393,8 @@ func TestEngine_UpdateNetworkMap(t *testing.T) {
 	}, MobileDependency{})
 
 	wgIface := &MockWGIface{
-		NameFunc: func() string { return "utun102" },
+		NameFunc:            func() string { return "utun102" },
+		IsUserspaceBindFunc: func() bool { return true },
 		RemovePeerFunc: func(peerKey string) error {
 			return nil
 		},

client/internal/getent/cgo_unix.go

@@ -0,0 +1,36 @@
+//go:build cgo && !osusergo && !windows
+
+package getent
+
+import "os/user"
+
+// Built with cgo, os/user resolves through libc (getpwnam_r and friends),
+// which goes through the host's NSS stack natively. Whatever it fails to
+// find, the getent command would not find either, so there is nothing to
+// fall back to.
+
+// LookupUser looks up a user by name.
+func LookupUser(username string) (*user.User, error) {
+	return user.Lookup(username)
+}
+
+// LookupUserID looks up a user by UID.
+func LookupUserID(uid string) (*user.User, error) {
+	return user.LookupId(uid)
+}
+
+// CurrentUser returns the user this process runs as.
+func CurrentUser() (*user.User, error) {
+	return user.Current()
+}
+
+// LookupGroupID looks up a group by GID.
+func LookupGroupID(gid string) (*user.Group, error) {
+	return user.LookupGroupId(gid)
+}
+
+// GroupIDs returns the IDs of the groups the user is a member of; libc's
+// getgrouplist handles NSS groups natively.
+func GroupIDs(u *user.User) ([]string, error) {
+	return u.GroupIds()
+}

client/internal/getent/getent.go

@@ -0,0 +1,6 @@
+// Package getent resolves users and groups through the host's NSS stack.
+// Built without cgo, os/user reads /etc/passwd and /etc/group alone and misses
+// anything LDAP, SSSD or winbind provide; the getent and id commands resolve
+// through NSS whatever the build. The lookups here try the standard library
+// first, which needs no subprocess, and fall back to those commands.
+package getent

client/internal/getent/getent_test.go

@@ -0,0 +1,155 @@
+package getent
+
+import (
+	"os/user"
+	"runtime"
+	"strconv"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestLookupUser_CurrentUser(t *testing.T) {
+	// The current user should always be resolvable on any platform
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	u, err := LookupUser(current.Username)
+	require.NoError(t, err)
+	assert.Equal(t, current.Username, u.Username)
+	assert.Equal(t, current.Uid, u.Uid)
+	assert.Equal(t, current.Gid, u.Gid)
+}
+
+func TestLookupUser_NonexistentUser(t *testing.T) {
+	_, err := LookupUser("nonexistent_user_xyzzy_12345")
+	require.Error(t, err, "should fail for nonexistent user")
+}
+
+func TestLookupUserID_CurrentUser(t *testing.T) {
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	u, err := LookupUserID(current.Uid)
+	require.NoError(t, err)
+	assert.Equal(t, current.Username, u.Username)
+	assert.Equal(t, current.Uid, u.Uid)
+}
+
+func TestCurrentUser(t *testing.T) {
+	stdUser, err := user.Current()
+	require.NoError(t, err)
+
+	u, err := CurrentUser()
+	require.NoError(t, err)
+	assert.Equal(t, stdUser.Uid, u.Uid)
+	assert.Equal(t, stdUser.Username, u.Username)
+}
+
+func TestGroupIDs_CurrentUser(t *testing.T) {
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	groups, err := GroupIDs(current)
+	require.NoError(t, err)
+	require.NotEmpty(t, groups, "current user should have at least one group")
+
+	if runtime.GOOS != "windows" {
+		for _, gid := range groups {
+			_, err := strconv.ParseUint(gid, 10, 32)
+			assert.NoError(t, err, "group ID %q should be a valid uint32", gid)
+		}
+	}
+}
+
+func TestUserShell_CurrentUser(t *testing.T) {
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	// getent may not be available on all systems (e.g., macOS without
+	// Homebrew getent), and Windows has no login shells at all.
+	shell, err := UserShell(current.Uid)
+	if err != nil {
+		t.Logf("UserShell failed, getent may not be available: %v", err)
+		return
+	}
+	if shell == "" {
+		t.Log("UserShell returned empty, the user has no shell set")
+		return
+	}
+	assert.True(t, shell[0] == '/', "shell should be an absolute path, got %q", shell)
+}
+
+func TestLookupUser_RootUser(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("no root user on Windows")
+	}
+
+	u, err := LookupUser("root")
+	if err != nil {
+		t.Skip("root user not available on this system")
+	}
+	assert.Equal(t, "0", u.Uid, "root should have UID 0")
+}
+
+// TestIntegration_FullLookupChain exercises the complete user lookup chain
+// against the real system, testing that all wrappers (LookupUser,
+// CurrentUser, GroupIDs, UserShell) produce consistent and correct results
+// when composed together.
+func TestIntegration_FullLookupChain(t *testing.T) {
+	// Step 1: CurrentUser must resolve the running user.
+	current, err := CurrentUser()
+	require.NoError(t, err, "CurrentUser must resolve the running user")
+	require.NotEmpty(t, current.Uid)
+	require.NotEmpty(t, current.Username)
+
+	// Step 2: LookupUser by the same username must return matching identity.
+	byName, err := LookupUser(current.Username)
+	require.NoError(t, err)
+	assert.Equal(t, current.Uid, byName.Uid, "lookup by name should return same UID")
+	assert.Equal(t, current.Gid, byName.Gid, "lookup by name should return same GID")
+	assert.Equal(t, current.HomeDir, byName.HomeDir, "lookup by name should return same home")
+
+	// Step 3: GroupIDs must return at least the primary GID.
+	groups, err := GroupIDs(current)
+	require.NoError(t, err)
+	require.NotEmpty(t, groups, "user must have at least one group")
+
+	foundPrimary := false
+	for _, gid := range groups {
+		if runtime.GOOS != "windows" {
+			_, err := strconv.ParseUint(gid, 10, 32)
+			require.NoError(t, err, "group ID %q must be a valid uint32", gid)
+		}
+		if gid == current.Gid {
+			foundPrimary = true
+		}
+	}
+	assert.True(t, foundPrimary, "primary GID %s should appear in supplementary groups", current.Gid)
+}
+
+// TestIntegration_LookupAndGroupsConsistency verifies that a user resolved via
+// LookupUser can have their groups resolved via GroupIDs, testing the handoff
+// between the two functions as used by the SSH server.
+func TestIntegration_LookupAndGroupsConsistency(t *testing.T) {
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	// Simulate the SSH server flow: lookup user, then get their groups.
+	resolved, err := LookupUser(current.Username)
+	require.NoError(t, err)
+
+	groups, err := GroupIDs(resolved)
+	require.NoError(t, err)
+	require.NotEmpty(t, groups, "resolved user must have groups")
+
+	// On Unix, all returned GIDs must be valid numeric values.
+	// On Windows, group IDs are SIDs (e.g., "S-1-5-32-544").
+	if runtime.GOOS != "windows" {
+		for _, gid := range groups {
+			_, err := strconv.ParseUint(gid, 10, 32)
+			assert.NoError(t, err, "group ID %q should be numeric", gid)
+		}
+	}
+}

client/internal/getent/nocgo_unix.go

@@ -0,0 +1,110 @@
+//go:build (!cgo || osusergo) && !windows
+
+package getent
+
+import (
+	"os"
+	"os/user"
+	"strconv"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// Without cgo, os/user only reads /etc/passwd and /etc/group and misses
+// NSS-provided users and groups; the getent and id commands go through the
+// host's NSS stack.
+
+// LookupUser looks up a user by name, falling back to getent if os/user fails.
+func LookupUser(username string) (*user.User, error) {
+	u, err := user.Lookup(username)
+	if err == nil {
+		return u, nil
+	}
+
+	stdErr := err
+	log.Debugf("os/user.Lookup(%q) failed, trying getent: %v", username, err)
+
+	u, _, getentErr := passwdLookup(username)
+	if getentErr != nil {
+		log.Debugf("getent fallback for %q also failed: %v", username, getentErr)
+		return nil, stdErr
+	}
+	return u, nil
+}
+
+// LookupUserID looks up a user by UID, falling back to getent if os/user fails.
+func LookupUserID(uid string) (*user.User, error) {
+	u, err := user.LookupId(uid)
+	if err == nil {
+		return u, nil
+	}
+
+	stdErr := err
+	log.Debugf("os/user.LookupId(%q) failed, trying getent: %v", uid, err)
+
+	u, _, getentErr := passwdLookup(uid)
+	if getentErr != nil {
+		log.Debugf("getent fallback for uid %s also failed: %v", uid, getentErr)
+		return nil, stdErr
+	}
+	return u, nil
+}
+
+// CurrentUser returns the user this process runs as, falling back to getent
+// if os/user fails.
+func CurrentUser() (*user.User, error) {
+	u, err := user.Current()
+	if err == nil {
+		return u, nil
+	}
+
+	stdErr := err
+	uid := strconv.Itoa(os.Getuid())
+	log.Debugf("os/user.Current() failed, trying getent with UID %s: %v", uid, err)
+
+	u, _, getentErr := passwdLookup(uid)
+	if getentErr != nil {
+		return nil, stdErr
+	}
+	return u, nil
+}
+
+// LookupGroupID looks up a group by GID, falling back to getent if os/user
+// fails.
+func LookupGroupID(gid string) (*user.Group, error) {
+	g, err := user.LookupGroupId(gid)
+	if err == nil {
+		return g, nil
+	}
+
+	stdErr := err
+	log.Debugf("os/user.LookupGroupId(%q) failed, trying getent: %v", gid, err)
+
+	g, _, getentErr := groupLookup(gid)
+	if getentErr != nil {
+		log.Debugf("getent fallback for gid %s also failed: %v", gid, getentErr)
+		return nil, stdErr
+	}
+	return g, nil
+}
+
+// GroupIDs returns the IDs of the groups the user is a member of.
+// NOTE: unlike the lookups above, which try the standard library first, this
+// intentionally tries `id -G` first because without cgo, user.GroupIds only
+// reads /etc/group and silently returns incomplete results for NSS users
+// (no error, just missing groups). The id command goes through NSS and
+// returns the full set.
+func GroupIDs(u *user.User) ([]string, error) {
+	ids, err := idGroups(u.Username)
+	if err == nil {
+		return ids, nil
+	}
+
+	log.Debugf("id -G %q failed, falling back to user.GroupIds(): %v", u.Username, err)
+
+	ids, stdErr := u.GroupIds()
+	if stdErr != nil {
+		return nil, stdErr
+	}
+	return ids, nil
+}

client/internal/getent/unix.go

@@ -0,0 +1,224 @@
+//go:build !windows
+
+package getent
+
+import (
+	"bufio"
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"os/user"
+	"runtime"
+	"strings"
+	"time"
+
+	log "github.com/sirupsen/logrus"
+)
+
+const commandTimeout = 5 * time.Second
+
+// groupFile lists which accounts are in which group, for hosts where the
+// getent command is not available (macOS ships without it).
+const groupFile = "/etc/group"
+
+// UserShell returns the login shell getent reports for the user with this UID.
+// It reaches shells that /etc/passwd does not list, because getent resolves
+// through the host's NSS stack.
+func UserShell(uid string) (string, error) {
+	_, shell, err := passwdLookup(uid)
+	if err != nil {
+		return "", err
+	}
+	return shell, nil
+}
+
+// GroupMembers returns the names of the group's members: from getent, which
+// resolves through NSS, or from /etc/group where getent is not available. A
+// group neither source describes is an error; an empty member list is not,
+// since accounts with the group as their primary one are not listed in it.
+func GroupMembers(name string) ([]string, error) {
+	_, members, err := groupLookup(name)
+	if err == nil {
+		return members, nil
+	}
+	log.Debugf("getent cannot list group %q, reading %s: %v", name, groupFile, err)
+	return groupMembersFromFile(groupFile, name)
+}
+
+// passwdLookup executes `getent passwd <query>`, where query is a username or
+// UID, and returns the user and login shell.
+func passwdLookup(query string) (*user.User, string, error) {
+	out, err := run("passwd", query)
+	if err != nil {
+		return nil, "", err
+	}
+	return parsePasswd(string(out))
+}
+
+// groupLookup executes `getent group <query>`, where query is a group name or
+// GID, and returns the group and its member names.
+func groupLookup(query string) (*user.Group, []string, error) {
+	out, err := run("group", query)
+	if err != nil {
+		return nil, nil, err
+	}
+	return parseGroup(string(out))
+}
+
+// run executes `getent <database> <key>` with a timeout.
+func run(database, key string) ([]byte, error) {
+	if !validateInput(key) {
+		return nil, fmt.Errorf("invalid getent input: %q", key)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), commandTimeout)
+	defer cancel()
+
+	out, err := exec.CommandContext(ctx, "getent", database, key).Output()
+	if err != nil {
+		return nil, fmt.Errorf("getent %s %s: %w", database, key, err)
+	}
+	return out, nil
+}
+
+// parsePasswd parses getent passwd output: "name:x:uid:gid:gecos:home:shell"
+func parsePasswd(output string) (*user.User, string, error) {
+	fields := strings.SplitN(strings.TrimSpace(output), ":", 8)
+	if len(fields) < 6 {
+		return nil, "", fmt.Errorf("unexpected getent output (need 6+ fields): %q", output)
+	}
+
+	if fields[0] == "" || fields[2] == "" || fields[3] == "" {
+		return nil, "", fmt.Errorf("missing required fields in getent output: %q", output)
+	}
+
+	var shell string
+	if len(fields) >= 7 {
+		shell = fields[6]
+	}
+
+	return &user.User{
+		Username: fields[0],
+		Uid:      fields[2],
+		Gid:      fields[3],
+		Name:     fields[4],
+		HomeDir:  fields[5],
+	}, shell, nil
+}
+
+// parseGroup parses getent group output: "name:x:gid:member,member"
+func parseGroup(output string) (*user.Group, []string, error) {
+	fields := strings.SplitN(strings.TrimSpace(output), ":", 4)
+	if len(fields) < 3 {
+		return nil, nil, fmt.Errorf("unexpected getent output (need 3+ fields): %q", output)
+	}
+
+	if fields[0] == "" || fields[2] == "" {
+		return nil, nil, fmt.Errorf("missing required fields in getent output: %q", output)
+	}
+
+	var members []string
+	if len(fields) >= 4 {
+		members = splitMembers(fields[3])
+	}
+	return &user.Group{Name: fields[0], Gid: fields[2]}, members, nil
+}
+
+func splitMembers(list string) []string {
+	var members []string
+	for member := range strings.SplitSeq(list, ",") {
+		if member != "" {
+			members = append(members, member)
+		}
+	}
+	return members
+}
+
+// groupMembersFromFile finds the group's member list in a file of /etc/group's
+// format. A group the file does not describe, because it comes from LDAP or
+// another NSS source, is an error rather than an empty list.
+func groupMembersFromFile(path, name string) ([]string, error) {
+	file, err := os.Open(path)
+	if err != nil {
+		return nil, fmt.Errorf("open %s: %w", path, err)
+	}
+	defer func() {
+		if err := file.Close(); err != nil {
+			log.Debugf("close %s: %v", path, err)
+		}
+	}()
+
+	scanner := bufio.NewScanner(file)
+	for scanner.Scan() {
+		// name:password:gid:member,member
+		fields := strings.Split(scanner.Text(), ":")
+		if len(fields) < 4 || fields[0] != name {
+			continue
+		}
+		return splitMembers(fields[3]), nil
+	}
+	if err := scanner.Err(); err != nil {
+		return nil, fmt.Errorf("read %s: %w", path, err)
+	}
+	return nil, fmt.Errorf("%s does not describe group %q", path, name)
+}
+
+// validateInput checks that the input is safe to pass to getent or id.
+// Allows POSIX usernames, numeric IDs, and common NSS extensions
+// (@ for Kerberos, $ for Samba, + for NIS compat). A leading hyphen is
+// rejected so the input can never be parsed as a command-line flag.
+func validateInput(input string) bool {
+	maxLen := 32
+	if runtime.GOOS == "linux" {
+		maxLen = 256
+	}
+
+	if len(input) == 0 || len(input) > maxLen {
+		return false
+	}
+
+	if input[0] == '-' {
+		return false
+	}
+
+	for _, r := range input {
+		if isAllowedChar(r) {
+			continue
+		}
+		return false
+	}
+	return true
+}
+
+func isAllowedChar(r rune) bool {
+	if r >= 'a' && r <= 'z' || r >= 'A' && r <= 'Z' || r >= '0' && r <= '9' {
+		return true
+	}
+	switch r {
+	case '.', '_', '-', '@', '+', '$':
+		return true
+	}
+	return false
+}
+
+// idGroups runs `id -G <username>` and returns the space-separated group IDs.
+func idGroups(username string) ([]string, error) {
+	if !validateInput(username) {
+		return nil, fmt.Errorf("invalid username for id command: %q", username)
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), commandTimeout)
+	defer cancel()
+
+	out, err := exec.CommandContext(ctx, "id", "-G", username).Output()
+	if err != nil {
+		return nil, fmt.Errorf("id -G %s: %w", username, err)
+	}
+
+	trimmed := strings.TrimSpace(string(out))
+	if trimmed == "" {
+		return nil, fmt.Errorf("id -G %s: empty output", username)
+	}
+	return strings.Fields(trimmed), nil
+}

client/internal/getent/unix_test.go

@@ -0,0 +1,471 @@
+//go:build !windows
+
+package getent
+
+import (
+	"os"
+	"os/exec"
+	"os/user"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+func TestParsePasswd(t *testing.T) {
+	tests := []struct {
+		name        string
+		input       string
+		wantUser    *user.User
+		wantShell   string
+		wantErr     bool
+		errContains string
+	}{
+		{
+			name:  "standard entry",
+			input: "alice:x:1001:1001:Alice Smith:/home/alice:/bin/bash\n",
+			wantUser: &user.User{
+				Username: "alice",
+				Uid:      "1001",
+				Gid:      "1001",
+				Name:     "Alice Smith",
+				HomeDir:  "/home/alice",
+			},
+			wantShell: "/bin/bash",
+		},
+		{
+			name:  "root entry",
+			input: "root:x:0:0:root:/root:/bin/bash",
+			wantUser: &user.User{
+				Username: "root",
+				Uid:      "0",
+				Gid:      "0",
+				Name:     "root",
+				HomeDir:  "/root",
+			},
+			wantShell: "/bin/bash",
+		},
+		{
+			name:  "empty gecos field",
+			input: "svc:x:999:999::/var/lib/svc:/usr/sbin/nologin",
+			wantUser: &user.User{
+				Username: "svc",
+				Uid:      "999",
+				Gid:      "999",
+				Name:     "",
+				HomeDir:  "/var/lib/svc",
+			},
+			wantShell: "/usr/sbin/nologin",
+		},
+		{
+			name:  "gecos with commas",
+			input: "john:x:1002:1002:John Doe,Room 101,555-1234,555-4321:/home/john:/bin/zsh",
+			wantUser: &user.User{
+				Username: "john",
+				Uid:      "1002",
+				Gid:      "1002",
+				Name:     "John Doe,Room 101,555-1234,555-4321",
+				HomeDir:  "/home/john",
+			},
+			wantShell: "/bin/zsh",
+		},
+		{
+			name:  "remote user with large UID",
+			input: "remoteuser:*:50001:50001:Remote User:/home/remoteuser:/bin/bash\n",
+			wantUser: &user.User{
+				Username: "remoteuser",
+				Uid:      "50001",
+				Gid:      "50001",
+				Name:     "Remote User",
+				HomeDir:  "/home/remoteuser",
+			},
+			wantShell: "/bin/bash",
+		},
+		{
+			name:  "no shell field (only 6 fields)",
+			input: "minimal:x:1000:1000::/home/minimal",
+			wantUser: &user.User{
+				Username: "minimal",
+				Uid:      "1000",
+				Gid:      "1000",
+				Name:     "",
+				HomeDir:  "/home/minimal",
+			},
+			wantShell: "",
+		},
+		{
+			name:        "too few fields",
+			input:       "bad:x:1000",
+			wantErr:     true,
+			errContains: "need 6+ fields",
+		},
+		{
+			name:        "empty username",
+			input:       ":x:1000:1000::/home/test:/bin/bash",
+			wantErr:     true,
+			errContains: "missing required fields",
+		},
+		{
+			name:        "empty UID",
+			input:       "test:x::1000::/home/test:/bin/bash",
+			wantErr:     true,
+			errContains: "missing required fields",
+		},
+		{
+			name:        "empty GID",
+			input:       "test:x:1000:::/home/test:/bin/bash",
+			wantErr:     true,
+			errContains: "missing required fields",
+		},
+		{
+			name:        "empty input",
+			input:       "",
+			wantErr:     true,
+			errContains: "need 6+ fields",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			u, shell, err := parsePasswd(tt.input)
+			if tt.wantErr {
+				require.Error(t, err)
+				if tt.errContains != "" {
+					assert.Contains(t, err.Error(), tt.errContains)
+				}
+				return
+			}
+			require.NoError(t, err)
+			assert.Equal(t, tt.wantUser.Username, u.Username, "username")
+			assert.Equal(t, tt.wantUser.Uid, u.Uid, "UID")
+			assert.Equal(t, tt.wantUser.Gid, u.Gid, "GID")
+			assert.Equal(t, tt.wantUser.Name, u.Name, "name/gecos")
+			assert.Equal(t, tt.wantUser.HomeDir, u.HomeDir, "home directory")
+			assert.Equal(t, tt.wantShell, shell, "shell")
+		})
+	}
+}
+
+func TestParseGroup(t *testing.T) {
+	tests := []struct {
+		name        string
+		input       string
+		wantGroup   *user.Group
+		wantMembers []string
+		wantErr     bool
+	}{
+		{
+			name:      "no members",
+			input:     "vma:x:1000:\n",
+			wantGroup: &user.Group{Name: "vma", Gid: "1000"},
+		},
+		{
+			name:        "one member",
+			input:       "sudo:x:27:alice",
+			wantGroup:   &user.Group{Name: "sudo", Gid: "27"},
+			wantMembers: []string{"alice"},
+		},
+		{
+			name:        "several members",
+			input:       "docker:x:998:alice,bob\n",
+			wantGroup:   &user.Group{Name: "docker", Gid: "998"},
+			wantMembers: []string{"alice", "bob"},
+		},
+		{
+			name:    "too few fields",
+			input:   "bad:x",
+			wantErr: true,
+		},
+		{
+			name:    "empty group name",
+			input:   ":x:1000:alice",
+			wantErr: true,
+		},
+		{
+			name:    "empty GID",
+			input:   "vma:x::alice",
+			wantErr: true,
+		},
+		{
+			name:    "empty input",
+			input:   "",
+			wantErr: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			g, members, err := parseGroup(tt.input)
+			if tt.wantErr {
+				require.Error(t, err)
+				return
+			}
+			require.NoError(t, err)
+			assert.Equal(t, tt.wantGroup.Name, g.Name, "group name")
+			assert.Equal(t, tt.wantGroup.Gid, g.Gid, "GID")
+			assert.Equal(t, tt.wantMembers, members, "members")
+		})
+	}
+}
+
+func TestGroupMembersFromFile(t *testing.T) {
+	tests := []struct {
+		name  string
+		entry string
+		want  []string
+	}{
+		{name: "no members", entry: "vma:x:1000:"},
+		{name: "only the owner", entry: "vma:x:1000:vma", want: []string{"vma"}},
+		{name: "two members", entry: "vma:x:1000:vma,bob", want: []string{"vma", "bob"}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			path := filepath.Join(t.TempDir(), "group")
+			body := "root:x:0:\n" + tt.entry + "\nsudo:x:27:vma\n"
+			require.NoError(t, os.WriteFile(path, []byte(body), 0o644), "write the group file")
+
+			members, err := groupMembersFromFile(path, "vma")
+			require.NoError(t, err, "entry %q", tt.entry)
+			assert.Equal(t, tt.want, members, "entry %q", tt.entry)
+		})
+	}
+}
+
+// A group the file does not describe, because it comes from LDAP or another
+// NSS source, is an error rather than an empty member list: the caller must
+// be able to tell "no members" from "no answer".
+func TestGroupMembersFromFileUnknownGroup(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "group")
+	require.NoError(t, os.WriteFile(path, []byte("root:x:0:\n"), 0o644), "write the group file")
+
+	_, err := groupMembersFromFile(path, "vma")
+	assert.Error(t, err, "a group the file does not describe")
+
+	_, err = groupMembersFromFile(filepath.Join(t.TempDir(), "absent"), "vma")
+	assert.Error(t, err, "no group file at all")
+}
+
+// GroupMembers on the root group, which every Unix has, whichever source
+// answers for it.
+func TestGroupMembers_RootGroup(t *testing.T) {
+	rootGroup := "root"
+	switch runtime.GOOS {
+	case "darwin", "dragonfly", "freebsd", "netbsd", "openbsd":
+		rootGroup = "wheel"
+	}
+
+	_, err := GroupMembers(rootGroup)
+	assert.NoError(t, err, "the %s group must be describable", rootGroup)
+}
+
+func TestValidateInput(t *testing.T) {
+	tests := []struct {
+		name  string
+		input string
+		want  bool
+	}{
+		{"normal username", "alice", true},
+		{"numeric UID", "1001", true},
+		{"dots and underscores", "alice.bob_test", true},
+		{"hyphen", "alice-bob", true},
+		{"leading hyphen rejected", "-i", false},
+		{"leading double hyphen rejected", "--no-idn", false},
+		{"lone hyphen rejected", "-", false},
+		{"kerberos principal", "user@REALM", true},
+		{"samba machine account", "MACHINE$", true},
+		{"NIS compat", "+user", true},
+		{"empty", "", false},
+		{"null byte", "alice\x00bob", false},
+		{"newline", "alice\nbob", false},
+		{"tab", "alice\tbob", false},
+		{"control char", "alice\x01bob", false},
+		{"DEL char", "alice\x7fbob", false},
+		{"space rejected", "alice bob", false},
+		{"semicolon rejected", "alice;bob", false},
+		{"backtick rejected", "alice`bob", false},
+		{"pipe rejected", "alice|bob", false},
+		{"33 chars exceeds non-linux max", makeLongString(33), runtime.GOOS == "linux"},
+		{"256 chars at linux max", makeLongString(256), runtime.GOOS == "linux"},
+		{"257 chars exceeds all limits", makeLongString(257), false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, validateInput(tt.input))
+		})
+	}
+}
+
+func makeLongString(n int) string {
+	b := make([]byte, n)
+	for i := range b {
+		b[i] = 'a'
+	}
+	return string(b)
+}
+
+func TestPasswdLookup_RootUser(t *testing.T) {
+	if _, err := exec.LookPath("getent"); err != nil {
+		t.Skip("getent not available on this system")
+	}
+
+	u, shell, err := passwdLookup("root")
+	require.NoError(t, err)
+	assert.Equal(t, "root", u.Username)
+	assert.Equal(t, "0", u.Uid)
+	assert.Equal(t, "0", u.Gid)
+	assert.NotEmpty(t, shell, "root should have a shell")
+}
+
+func TestPasswdLookup_ByUID(t *testing.T) {
+	if _, err := exec.LookPath("getent"); err != nil {
+		t.Skip("getent not available on this system")
+	}
+
+	u, _, err := passwdLookup("0")
+	require.NoError(t, err)
+	assert.Equal(t, "root", u.Username)
+	assert.Equal(t, "0", u.Uid)
+}
+
+func TestPasswdLookup_NonexistentUser(t *testing.T) {
+	if _, err := exec.LookPath("getent"); err != nil {
+		t.Skip("getent not available on this system")
+	}
+
+	_, _, err := passwdLookup("nonexistent_user_xyzzy_12345")
+	assert.Error(t, err)
+}
+
+func TestPasswdLookup_InvalidInput(t *testing.T) {
+	_, _, err := passwdLookup("")
+	assert.Error(t, err)
+
+	_, _, err = passwdLookup("user\x00name")
+	assert.Error(t, err)
+}
+
+func TestPasswdLookup_NotAvailable(t *testing.T) {
+	if _, err := exec.LookPath("getent"); err == nil {
+		t.Skip("getent is available, can't test missing case")
+	}
+
+	_, _, err := passwdLookup("root")
+	assert.Error(t, err, "should fail when getent is not installed")
+}
+
+func TestGroupLookup_RootGroup(t *testing.T) {
+	if _, err := exec.LookPath("getent"); err != nil {
+		t.Skip("getent not available on this system")
+	}
+
+	g, _, err := groupLookup("0")
+	require.NoError(t, err)
+	assert.Equal(t, "0", g.Gid, "GID 0 resolves to the root group")
+	assert.NotEmpty(t, g.Name, "the root group has a name")
+}
+
+func TestIdGroups_CurrentUser(t *testing.T) {
+	if _, err := exec.LookPath("id"); err != nil {
+		t.Skip("id not available on this system")
+	}
+
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	groups, err := idGroups(current.Username)
+	require.NoError(t, err)
+	require.NotEmpty(t, groups, "current user should have at least one group")
+
+	for _, gid := range groups {
+		_, err := strconv.ParseUint(gid, 10, 32)
+		assert.NoError(t, err, "group ID %q should be a valid uint32", gid)
+	}
+}
+
+func TestIdGroups_NonexistentUser(t *testing.T) {
+	if _, err := exec.LookPath("id"); err != nil {
+		t.Skip("id not available on this system")
+	}
+
+	_, err := idGroups("nonexistent_user_xyzzy_12345")
+	assert.Error(t, err)
+}
+
+func TestIdGroups_InvalidInput(t *testing.T) {
+	_, err := idGroups("")
+	assert.Error(t, err)
+
+	_, err = idGroups("user\x00name")
+	assert.Error(t, err)
+}
+
+func TestGetentResultsMatchStdlib(t *testing.T) {
+	if _, err := exec.LookPath("getent"); err != nil {
+		t.Skip("getent not available on this system")
+	}
+
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	getentUser, _, err := passwdLookup(current.Username)
+	require.NoError(t, err)
+
+	assert.Equal(t, current.Username, getentUser.Username, "username should match")
+	assert.Equal(t, current.Uid, getentUser.Uid, "UID should match")
+	assert.Equal(t, current.Gid, getentUser.Gid, "GID should match")
+	assert.Equal(t, current.HomeDir, getentUser.HomeDir, "home directory should match")
+}
+
+func TestGetentResultsMatchStdlib_ByUID(t *testing.T) {
+	if _, err := exec.LookPath("getent"); err != nil {
+		t.Skip("getent not available on this system")
+	}
+
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	getentUser, _, err := passwdLookup(current.Uid)
+	require.NoError(t, err)
+
+	assert.Equal(t, current.Username, getentUser.Username, "username should match when looked up by UID")
+	assert.Equal(t, current.Uid, getentUser.Uid, "UID should match")
+}
+
+func TestIdGroupsMatchStdlib(t *testing.T) {
+	if _, err := exec.LookPath("id"); err != nil {
+		t.Skip("id not available on this system")
+	}
+
+	current, err := user.Current()
+	require.NoError(t, err)
+
+	stdGroups, err := current.GroupIds()
+	if err != nil {
+		t.Skip("os/user.GroupIds() not working, likely CGO_ENABLED=0")
+	}
+
+	idGroupIDs, err := idGroups(current.Username)
+	require.NoError(t, err)
+
+	// Deduplicate both lists: id -G can return duplicates (e.g., root in Docker)
+	// and ElementsMatch treats duplicates as distinct.
+	assert.ElementsMatch(t, uniqueStrings(stdGroups), uniqueStrings(idGroupIDs), "id -G should return same groups as os/user")
+}
+
+func uniqueStrings(ss []string) []string {
+	seen := make(map[string]struct{}, len(ss))
+	out := make([]string, 0, len(ss))
+	for _, s := range ss {
+		if _, ok := seen[s]; ok {
+			continue
+		}
+		seen[s] = struct{}{}
+		out = append(out, s)
+	}
+	return out
+}

client/internal/getent/windows.go

@@ -0,0 +1,36 @@
+//go:build windows
+
+package getent
+
+import (
+	"errors"
+	"os/user"
+)
+
+// Windows does not use NSS or getent; os/user resolves accounts there
+// without cgo, so everything delegates to it.
+
+// LookupUser looks up a user by name.
+func LookupUser(username string) (*user.User, error) {
+	return user.Lookup(username)
+}
+
+// LookupUserID looks up a user by UID.
+func LookupUserID(uid string) (*user.User, error) {
+	return user.LookupId(uid)
+}
+
+// CurrentUser returns the user this process runs as.
+func CurrentUser() (*user.User, error) {
+	return user.Current()
+}
+
+// GroupIDs returns the IDs of the groups the user is a member of.
+func GroupIDs(u *user.User) ([]string, error) {
+	return u.GroupIds()
+}
+
+// UserShell is unanswerable on Windows, which has no login-shell database.
+func UserShell(string) (string, error) {
+	return "", errors.ErrUnsupported
+}

client/internal/ingressgw/manager.go

@@ -24,14 +24,14 @@ type RulePair struct {
 type Manager struct {
 	dnatFirewall DNATFirewall
 
-	rules   map[string]RulePair // keys is the ID of the ForwardRule
+	rules   map[firewall.RuleID]RulePair
 	rulesMu sync.Mutex
 }
 
 func NewManager(dnatFirewall DNATFirewall) *Manager {
 	return &Manager{
 		dnatFirewall: dnatFirewall,
-		rules:        make(map[string]RulePair),
+		rules:        make(map[firewall.RuleID]RulePair),
 	}
 }
 
@@ -41,7 +41,7 @@ func (h *Manager) Update(forwardRules []firewall.ForwardRule) error {
 
 	var mErr *multierror.Error
 
-	toDelete := make(map[string]RulePair, len(h.rules))
+	toDelete := make(map[firewall.RuleID]RulePair, len(h.rules))
 	for id, r := range h.rules {
 		toDelete[id] = r
 	}
@@ -59,6 +59,10 @@ func (h *Manager) Update(forwardRules []firewall.ForwardRule) error {
 			mErr = multierror.Append(mErr, fmt.Errorf("add forward rule '%s': %v", fwdRule.String(), err))
 			continue
 		}
+		if rule == nil {
+			mErr = multierror.Append(mErr, fmt.Errorf("add forward rule '%s': backend returned no rule", fwdRule.String()))
+			continue
+		}
 		log.Infof("forward rule has been added '%s'", fwdRule)
 		h.rules[id] = RulePair{
 			ForwardRule: fwdRule,
@@ -90,7 +94,7 @@ func (h *Manager) Close() error {
 		}
 	}
 
-	h.rules = make(map[string]RulePair)
+	h.rules = make(map[firewall.RuleID]RulePair)
 	return nberrors.FormatErrorOrNil(mErr)
 }
 

client/internal/ingressgw/manager_test.go

@@ -14,11 +14,11 @@ var (
 )
 
 type MocFwRule struct {
-	id string
+	id firewall.RuleID
 }
 
-func (m *MocFwRule) ID() string {
-	return string(m.id)
+func (m *MocFwRule) ID() firewall.RuleID {
+	return m.id
 }
 
 type MockDNATFirewall struct {

client/internal/ipcauth/identity.go

@@ -91,6 +91,19 @@ func (i Identity) IsPrivileged() bool {
 	return slices.Contains(i.Groups, sidAdministrators)
 }
 
+// SameUser reports whether two identities are the same local principal. Only
+// the account is compared: the group set and the elevation flag describe what a
+// token may do, not who it belongs to. A SID on either side decides the
+// comparison, so a Windows principal never matches a Unix one on the UID both
+// happen to leave at zero. The zero Identity carries uid 0, so callers must
+// establish that both identities are real before the answer means anything.
+func (i Identity) SameUser(other Identity) bool {
+	if i.SID != "" || other.SID != "" {
+		return i.SID == other.SID
+	}
+	return i.UID == other.UID
+}
+
 // String renders the identity for audit logs and denial messages.
 func (i Identity) String() string {
 	if i.IsWindows() {

client/internal/ipcauth/identity_sameuser_test.go

@@ -0,0 +1,66 @@
+package ipcauth
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+func TestIdentitySameUser(t *testing.T) {
+	tests := []struct {
+		name string
+		a    Identity
+		b    Identity
+		want bool
+	}{
+		{
+			name: "same uid",
+			a:    Identity{UID: 1000, GID: 1000},
+			b:    Identity{UID: 1000, GID: 1000},
+			want: true,
+		},
+		{
+			name: "same uid, different gid and pid still the same user",
+			a:    Identity{UID: 1000, GID: 1000, PID: 11},
+			b:    Identity{UID: 1000, GID: 27, PID: 22},
+			want: true,
+		},
+		{
+			name: "different uid",
+			a:    Identity{UID: 1000},
+			b:    Identity{UID: 1001},
+			want: false,
+		},
+		{
+			name: "same sid",
+			a:    Identity{SID: "S-1-5-21-1-2-3-1001"},
+			b:    Identity{SID: "S-1-5-21-1-2-3-1001"},
+			want: true,
+		},
+		{
+			name: "same sid, elevation and groups differ",
+			a:    Identity{SID: "S-1-5-21-1-2-3-1001", Elevated: true, Groups: []string{sidAdministrators}},
+			b:    Identity{SID: "S-1-5-21-1-2-3-1001"},
+			want: true,
+		},
+		{
+			name: "different sid",
+			a:    Identity{SID: "S-1-5-21-1-2-3-1001"},
+			b:    Identity{SID: "S-1-5-21-1-2-3-1002"},
+			want: false,
+		},
+		{
+			name: "a windows principal is never a unix one",
+			a:    Identity{SID: "S-1-5-18"},
+			b:    Identity{UID: 0},
+			want: false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			assert.Equal(t, tt.want, tt.a.SameUser(tt.b))
+			assert.Equal(t, tt.want, tt.b.SameUser(tt.a), "SameUser must be symmetric")
+		})
+	}
+}

client/internal/ipcauth/privileged.go

@@ -91,6 +91,12 @@ func SelfDelegatesTo() (Identity, bool) {
 	return selfIdentity, true
 }
 
+// The values PrivilegedActorKey returns.
+const (
+	ActorKeyAdministrator = "administrator"
+	ActorKeyRoot          = "root"
+)
+
 // PrivilegedActor names the principal a privileged operation requires, for use
 // in messages shown to the user.
 func PrivilegedActor() string {
@@ -100,6 +106,16 @@ func PrivilegedActor() string {
 	return "root"
 }
 
+// PrivilegedActorKey identifies that principal without wording it, for a client
+// that writes its own message in the user's language. The words PrivilegedActor
+// returns are English, and a translated sentence cannot borrow them.
+func PrivilegedActorKey() string {
+	if runtime.GOOS == "windows" {
+		return ActorKeyAdministrator
+	}
+	return ActorKeyRoot
+}
+
 // ElevatedCommand renders a command so that running it grants the privileges the
 // operation needs. Windows has no in-line equivalent of sudo, so the command is
 // returned unchanged and the user is expected to run it from an elevated

client/internal/localmetrics/localmetrics.go

@@ -0,0 +1,274 @@
+// Package localmetrics exposes client connection state as a local
+// Prometheus /metrics endpoint.
+package localmetrics
+
+import (
+	"context"
+	"errors"
+	"net"
+	"net/http"
+	"net/netip"
+	"sync"
+	"time"
+
+	"github.com/prometheus/client_golang/prometheus"
+	"github.com/prometheus/client_golang/prometheus/promhttp"
+	dto "github.com/prometheus/client_model/go"
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/peer"
+)
+
+// DefaultListenAddress is used when local metrics are enabled without an explicit address.
+const DefaultListenAddress = "127.0.0.1:9191"
+
+const (
+	shutdownTimeout   = 3 * time.Second
+	readHeaderTimeout = 5 * time.Second
+	readTimeout       = 10 * time.Second
+	writeTimeout      = 30 * time.Second
+	idleTimeout       = time.Minute
+)
+
+// statusSource provides the connection state snapshots the collector reads on scrape.
+type statusSource interface {
+	GetPeerStates() []peer.State
+	GetManagementState() peer.ManagementState
+	GetSignalState() peer.SignalState
+}
+
+// GathererProvider returns the current client metrics gatherer, or nil when
+// no engine is running. It is called on every scrape.
+type GathererProvider func() prometheus.Gatherer
+
+// Manager runs the local /metrics HTTP endpoint according to the active
+// client configuration. Reconcile is safe to call on every config change.
+type Manager struct {
+	status        statusSource
+	clientMetrics GathererProvider
+
+	mu   sync.Mutex
+	srv  *http.Server
+	addr string
+}
+
+// NewManager creates a manager that serves metrics from status and
+// clientMetrics and shuts down when ctx is canceled.
+func NewManager(ctx context.Context, status statusSource, clientMetrics GathererProvider) *Manager {
+	m := &Manager{status: status, clientMetrics: clientMetrics}
+	go func() {
+		<-ctx.Done()
+		m.Stop()
+	}()
+	return m
+}
+
+// Reconcile starts, stops, or restarts the metrics endpoint to match the
+// desired state. An empty addr falls back to DefaultListenAddress.
+func (m *Manager) Reconcile(enabled bool, addr string) {
+	if addr == "" {
+		addr = DefaultListenAddress
+	}
+	warnIfNotLoopback(addr)
+
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	if !enabled {
+		m.stop()
+		return
+	}
+	if m.srv != nil && m.addr == addr {
+		return
+	}
+	m.stop()
+
+	registry := prometheus.NewRegistry()
+	registry.MustRegister(newCollector(m.status))
+
+	gatherers := prometheus.Gatherers{registry, prometheus.GathererFunc(func() ([]*dto.MetricFamily, error) {
+		if m.clientMetrics == nil {
+			return nil, nil
+		}
+		g := m.clientMetrics()
+		if g == nil {
+			return nil, nil
+		}
+		return g.Gather()
+	})}
+
+	mux := http.NewServeMux()
+	mux.Handle("/metrics", promhttp.HandlerFor(gatherers, promhttp.HandlerOpts{}))
+
+	srv := &http.Server{
+		Addr:              addr,
+		Handler:           mux,
+		ReadHeaderTimeout: readHeaderTimeout,
+		ReadTimeout:       readTimeout,
+		WriteTimeout:      writeTimeout,
+		IdleTimeout:       idleTimeout,
+	}
+	m.srv = srv
+	m.addr = addr
+
+	log.Infof("serving local metrics on http://%s/metrics", addr)
+	go func() {
+		if err := srv.ListenAndServe(); err != nil && !errors.Is(err, http.ErrServerClosed) {
+			log.Errorf("failed to serve local metrics on %s: %v", addr, err)
+			m.clear(srv)
+		}
+	}()
+}
+
+// clear drops the reference to srv so a later Reconcile with the same
+// address restarts it. A newer server may already have replaced it, in
+// which case the reference must stay.
+func (m *Manager) clear(srv *http.Server) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	if m.srv != srv {
+		return
+	}
+	m.srv = nil
+	m.addr = ""
+}
+
+// Stop shuts down the metrics endpoint if it is running.
+func (m *Manager) Stop() {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	m.stop()
+}
+
+// stop shuts down the running server. Callers must hold m.mu.
+func (m *Manager) stop() {
+	if m.srv == nil {
+		return
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), shutdownTimeout)
+	defer cancel()
+	if err := m.srv.Shutdown(ctx); err != nil {
+		log.Debugf("failed to shut down local metrics server: %v", err)
+	}
+	m.srv = nil
+	m.addr = ""
+}
+
+// collector converts status recorder snapshots into Prometheus metrics at scrape time.
+type collector struct {
+	status statusSource
+
+	managementConnected *prometheus.Desc
+	signalConnected     *prometheus.Desc
+	peersTotal          *prometheus.Desc
+	peersConnected      *prometheus.Desc
+	peerLatency         *prometheus.Desc
+}
+
+func newCollector(status statusSource) *collector {
+	return &collector{
+		status: status,
+		managementConnected: prometheus.NewDesc(
+			"netbird_management_connected",
+			"Whether the client is connected to the management service (1 connected, 0 disconnected).",
+			nil, nil,
+		),
+		signalConnected: prometheus.NewDesc(
+			"netbird_signal_connected",
+			"Whether the client is connected to the signal service (1 connected, 0 disconnected).",
+			nil, nil,
+		),
+		peersTotal: prometheus.NewDesc(
+			"netbird_peers",
+			"Number of peers known to this client.",
+			nil, nil,
+		),
+		peersConnected: prometheus.NewDesc(
+			"netbird_peers_connected",
+			"Number of connected peers by connection type.",
+			[]string{"connection_type"}, nil,
+		),
+		peerLatency: prometheus.NewDesc(
+			"netbird_peer_latency_seconds",
+			"Round-trip latency per directly connected peer; relayed connections have no latency measurement.",
+			[]string{"peer"}, nil,
+		),
+	}
+}
+
+// Describe implements prometheus.Collector.
+func (c *collector) Describe(ch chan<- *prometheus.Desc) {
+	ch <- c.managementConnected
+	ch <- c.signalConnected
+	ch <- c.peersTotal
+	ch <- c.peersConnected
+	ch <- c.peerLatency
+}
+
+// Collect implements prometheus.Collector.
+func (c *collector) Collect(ch chan<- prometheus.Metric) {
+	ch <- prometheus.MustNewConstMetric(c.managementConnected, prometheus.GaugeValue, boolToFloat(c.status.GetManagementState().Connected))
+	ch <- prometheus.MustNewConstMetric(c.signalConnected, prometheus.GaugeValue, boolToFloat(c.status.GetSignalState().Connected))
+
+	peers := c.status.GetPeerStates()
+	ch <- prometheus.MustNewConstMetric(c.peersTotal, prometheus.GaugeValue, float64(len(peers)))
+
+	var p2p, relayed float64
+	for _, p := range peers {
+		if p.ConnStatus != peer.StatusConnected {
+			continue
+		}
+		if p.Relayed {
+			relayed++
+			continue
+		}
+		p2p++
+
+		if latency := p.Latency.Seconds(); latency > 0 {
+			ch <- prometheus.MustNewConstMetric(c.peerLatency, prometheus.GaugeValue, latency, p.FQDN)
+		}
+	}
+	ch <- prometheus.MustNewConstMetric(c.peersConnected, prometheus.GaugeValue, p2p, "p2p")
+	ch <- prometheus.MustNewConstMetric(c.peersConnected, prometheus.GaugeValue, relayed, "relay")
+}
+
+func boolToFloat(b bool) float64 {
+	if b {
+		return 1
+	}
+	return 0
+}
+
+// IsLoopback reports whether addr binds the endpoint to the local host only.
+// An empty address means DefaultListenAddress. It fails closed: an address
+// that cannot be confirmed loopback, including an unparseable one, is not.
+func IsLoopback(addr string) bool {
+	if addr == "" {
+		addr = DefaultListenAddress
+	}
+
+	host, _, err := net.SplitHostPort(addr)
+	if err != nil {
+		return false
+	}
+	if host == "localhost" {
+		return true
+	}
+
+	ip, err := netip.ParseAddr(host)
+	if err != nil {
+		return false
+	}
+	return ip.Unmap().IsLoopback()
+}
+
+// warnIfNotLoopback logs a warning when the listen address cannot be
+// confirmed to be local-only, since the endpoint exposes peer and
+// connectivity details without authentication.
+func warnIfNotLoopback(addr string) {
+	if IsLoopback(addr) {
+		return
+	}
+	log.Warnf("local metrics endpoint listens on non-loopback address %s and is reachable from the network without authentication", addr)
+}

client/internal/localmetrics/localmetrics_test.go

@@ -0,0 +1,151 @@
+package localmetrics
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/prometheus/client_golang/prometheus/testutil"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/peer"
+)
+
+type stubStatus struct {
+	peers      []peer.State
+	management peer.ManagementState
+	signal     peer.SignalState
+}
+
+func (s *stubStatus) GetPeerStates() []peer.State              { return s.peers }
+func (s *stubStatus) GetManagementState() peer.ManagementState { return s.management }
+func (s *stubStatus) GetSignalState() peer.SignalState         { return s.signal }
+
+func testStatus() *stubStatus {
+	return &stubStatus{
+		management: peer.ManagementState{Connected: true},
+		signal:     peer.SignalState{Connected: true},
+		peers: []peer.State{
+			{FQDN: "peer-a.netbird.cloud", IP: "100.90.0.1", ConnStatus: peer.StatusConnected, Relayed: false, Latency: 12 * time.Millisecond},
+			{FQDN: "peer-b.netbird.cloud", IP: "100.90.0.2", ConnStatus: peer.StatusConnected, Relayed: false, Latency: 36 * time.Millisecond},
+			{FQDN: "peer-c.netbird.cloud", IP: "100.90.0.3", ConnStatus: peer.StatusConnected, Relayed: true},
+			{FQDN: "peer-d.netbird.cloud", IP: "100.90.0.4", ConnStatus: peer.StatusIdle},
+		},
+	}
+}
+
+func TestCollector(t *testing.T) {
+	c := newCollector(testStatus())
+
+	expected := `
+# HELP netbird_management_connected Whether the client is connected to the management service (1 connected, 0 disconnected).
+# TYPE netbird_management_connected gauge
+netbird_management_connected 1
+# HELP netbird_peer_latency_seconds Round-trip latency per directly connected peer; relayed connections have no latency measurement.
+# TYPE netbird_peer_latency_seconds gauge
+netbird_peer_latency_seconds{peer="peer-a.netbird.cloud"} 0.012
+netbird_peer_latency_seconds{peer="peer-b.netbird.cloud"} 0.036
+# HELP netbird_peers Number of peers known to this client.
+# TYPE netbird_peers gauge
+netbird_peers 4
+# HELP netbird_peers_connected Number of connected peers by connection type.
+# TYPE netbird_peers_connected gauge
+netbird_peers_connected{connection_type="p2p"} 2
+netbird_peers_connected{connection_type="relay"} 1
+# HELP netbird_signal_connected Whether the client is connected to the signal service (1 connected, 0 disconnected).
+# TYPE netbird_signal_connected gauge
+netbird_signal_connected 1
+`
+	require.NoError(t, testutil.CollectAndCompare(c, strings.NewReader(expected)))
+}
+
+func TestServe(t *testing.T) {
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	require.NoError(t, err, "must find a free port")
+	addr := ln.Addr().String()
+	require.NoError(t, ln.Close())
+
+	ctx, cancel := context.WithCancel(context.Background())
+	t.Cleanup(cancel)
+	m := NewManager(ctx, testStatus(), nil)
+	m.Reconcile(true, addr)
+
+	var body string
+	require.Eventually(t, func() bool {
+		resp, err := http.Get(fmt.Sprintf("http://%s/metrics", addr))
+		if err != nil {
+			return false
+		}
+		defer resp.Body.Close()
+		data, err := io.ReadAll(resp.Body)
+		if err != nil || resp.StatusCode != http.StatusOK {
+			return false
+		}
+		body = string(data)
+		return true
+	}, 2*time.Second, 50*time.Millisecond, "metrics endpoint should come up")
+
+	assert.Contains(t, body, "netbird_peers 4")
+	assert.Contains(t, body, `netbird_peers_connected{connection_type="relay"} 1`)
+	assert.Contains(t, body, `netbird_peer_latency_seconds{peer="peer-a.netbird.cloud"} 0.012`)
+}
+
+// A server that never came up must not be remembered, otherwise reconciling the
+// same address again is a no-op and the endpoint never recovers.
+func TestReconcileForgetsAFailedServer(t *testing.T) {
+	blocker, err := net.Listen("tcp", "127.0.0.1:0")
+	require.NoError(t, err, "must find a free port")
+	t.Cleanup(func() { _ = blocker.Close() })
+	addr := blocker.Addr().String()
+
+	ctx, cancel := context.WithCancel(context.Background())
+	t.Cleanup(cancel)
+	m := NewManager(ctx, testStatus(), nil)
+	m.Reconcile(true, addr)
+
+	require.Eventually(t, func() bool {
+		m.mu.Lock()
+		defer m.mu.Unlock()
+		return m.srv == nil && m.addr == ""
+	}, 2*time.Second, 20*time.Millisecond, "the failed server should be dropped")
+
+	require.NoError(t, blocker.Close())
+	m.Reconcile(true, addr)
+
+	require.Eventually(t, func() bool {
+		resp, err := http.Get(fmt.Sprintf("http://%s/metrics", addr))
+		if err != nil {
+			return false
+		}
+		defer resp.Body.Close()
+		return resp.StatusCode == http.StatusOK
+	}, 2*time.Second, 50*time.Millisecond, "reconciling the same address should retry the bind")
+}
+
+func TestIsLoopback(t *testing.T) {
+	tests := map[string]bool{
+		"":                        true,
+		"127.0.0.1:9191":          true,
+		"127.9.9.9:9191":          true,
+		"[::1]:9191":              true,
+		"[::ffff:127.0.0.1]:9191": true,
+		"localhost:9191":          true,
+		"0.0.0.0:9191":            false,
+		"[::]:9191":               false,
+		"192.168.1.10:9191":       false,
+		"not-an-address":          false,
+		"example.com:9191":        false,
+	}
+
+	for addr, want := range tests {
+		t.Run(addr, func(t *testing.T) {
+			assert.Equal(t, want, IsLoopback(addr), "loopback verdict for %q", addr)
+		})
+	}
+}

client/internal/message_convert.go

@@ -10,21 +10,6 @@ import (
 	mgmProto "github.com/netbirdio/netbird/shared/management/proto"
 )
 
-func convertToFirewallProtocol(protocol mgmProto.RuleProtocol) (firewallManager.Protocol, error) {
-	switch protocol {
-	case mgmProto.RuleProtocol_TCP:
-		return firewallManager.ProtocolTCP, nil
-	case mgmProto.RuleProtocol_UDP:
-		return firewallManager.ProtocolUDP, nil
-	case mgmProto.RuleProtocol_ICMP:
-		return firewallManager.ProtocolICMP, nil
-	case mgmProto.RuleProtocol_ALL:
-		return firewallManager.ProtocolALL, nil
-	default:
-		return "", fmt.Errorf("invalid protocol type: %s", protocol.String())
-	}
-}
-
 func convertPortInfo(portInfo *mgmProto.PortInfo) (*firewallManager.Port, error) {
 	if portInfo == nil {
 		return nil, errors.New("portInfo cannot be nil")

client/internal/metrics/connection_type.go

@@ -4,11 +4,17 @@ package metrics
 type ConnectionType string
 
 const (
-	// ConnectionTypeICE represents a direct peer-to-peer connection using ICE
-	ConnectionTypeICE ConnectionType = "ice"
+	// ConnectionTypeICEP2P represents a direct peer-to-peer connection using ICE
+	ConnectionTypeICEP2P ConnectionType = "ice_p2p"
+
+	// ConnectionTypeICETurn represents an ICE connection through a TURN server
+	ConnectionTypeICETurn ConnectionType = "ice_turn"
 
 	// ConnectionTypeRelay represents a relayed connection
 	ConnectionTypeRelay ConnectionType = "relay"
+
+	// ConnectionTypeUnknown represents a connection with no active transport. It is not pushed.
+	ConnectionTypeUnknown ConnectionType = "unknown"
 )
 
 // String returns the string representation of the connection type

client/internal/metrics/influxdb.go

@@ -45,30 +45,13 @@ func (m *influxDBMetrics) RecordConnectionStages(
 	isReconnection bool,
 	timestamps ConnectionStageTimestamps,
 ) {
-	var signalingReceivedToConnection, connectionToWgHandshake, totalDuration float64
-
-	if !timestamps.SignalingReceived.IsZero() && !timestamps.ConnectionReady.IsZero() {
-		signalingReceivedToConnection = timestamps.ConnectionReady.Sub(timestamps.SignalingReceived).Seconds()
-	}
-
-	if !timestamps.ConnectionReady.IsZero() && !timestamps.WgHandshakeSuccess.IsZero() {
-		connectionToWgHandshake = timestamps.WgHandshakeSuccess.Sub(timestamps.ConnectionReady).Seconds()
-	}
-
-	if !timestamps.SignalingReceived.IsZero() && !timestamps.WgHandshakeSuccess.IsZero() {
-		totalDuration = timestamps.WgHandshakeSuccess.Sub(timestamps.SignalingReceived).Seconds()
-	}
-
-	attemptType := "initial"
-	if isReconnection {
-		attemptType = "reconnection"
-	}
+	signalingReceivedToConnection, connectionToWgHandshake, totalDuration := timestamps.Durations()
 
 	connTypeStr := connectionType.String()
 	tags := fmt.Sprintf("deployment_type=%s,connection_type=%s,attempt_type=%s,version=%s,os=%s,arch=%s,peer_id=%s,connection_pair_id=%s",
 		agentInfo.DeploymentType.String(),
 		connTypeStr,
-		attemptType,
+		attemptType(isReconnection),
 		agentInfo.Version,
 		agentInfo.OS,
 		agentInfo.Arch,
@@ -94,7 +77,7 @@ func (m *influxDBMetrics) RecordConnectionStages(
 	m.trimLocked()
 
 	log.Tracef("peer connection metrics [%s, %s, %s]: signalingReceived→connection: %.3fs, connection→wg_handshake: %.3fs, total: %.3fs",
-		agentInfo.DeploymentType.String(), connTypeStr, attemptType, signalingReceivedToConnection, connectionToWgHandshake, totalDuration)
+		agentInfo.DeploymentType.String(), connTypeStr, attemptType(isReconnection), signalingReceivedToConnection, connectionToWgHandshake, totalDuration)
 }
 
 func (m *influxDBMetrics) RecordSyncDuration(_ context.Context, agentInfo AgentInfo, duration time.Duration) {

client/internal/metrics/influxdb_test.go

@@ -28,7 +28,7 @@ func TestInfluxDBMetrics_RecordAndExport(t *testing.T) {
 		WgHandshakeSuccess: time.Now().Add(-1 * time.Second),
 	}
 
-	m.RecordConnectionStages(context.Background(), agentInfo, "pair123", ConnectionTypeICE, false, ts)
+	m.RecordConnectionStages(context.Background(), agentInfo, "pair123", ConnectionTypeICEP2P, false, ts)
 
 	var buf bytes.Buffer
 	err := m.Export(&buf)
@@ -60,7 +60,7 @@ func TestInfluxDBMetrics_ExportDeterministicFieldOrder(t *testing.T) {
 
 	// Record multiple times and verify consistent field order
 	for i := 0; i < 10; i++ {
-		m.RecordConnectionStages(context.Background(), agentInfo, "pair123", ConnectionTypeICE, false, ts)
+		m.RecordConnectionStages(context.Background(), agentInfo, "pair123", ConnectionTypeICEP2P, false, ts)
 	}
 
 	var buf bytes.Buffer

client/internal/metrics/infra/README.md

@@ -32,13 +32,24 @@ Clients do not talk to InfluxDB directly. An ingest server sits between clients
 ```text
 Client ──POST──▶ Ingest Server (:8087) ──▶ InfluxDB (internal)
                   │
+                  ├─ Checks the X-Peer-ID header format
                   ├─ Validates line protocol
                   ├─ Allowlists measurements, fields, and tags
                   ├─ Rejects out-of-bound values
                   └─ Serves remote config at /config
 ```
 
-- **No secret/token-based client auth** — the ingest server holds the InfluxDB token server-side. Clients must send a hashed peer ID via `X-Peer-ID` header.
+- **Intentionally unauthenticated** — the endpoint receives obfuscated telemetry from
+  the peers of both cloud and self-hosted deployments. For a self-hosted peer there is
+  no shared trust anchor with this server, so there is nothing to authenticate against.
+- **`X-Peer-ID` is a correlation tag, not a credential** — it carries the obfuscated
+  peer identifier so samples from one peer can be grouped. The server only checks that
+  the header is well-formed (16 hex chars); a malformed value is rejected with
+  `400 Bad Request`, not `401`. Any well-formed value is accepted by design, and the
+  header must not be relied on for access control. The header itself is not forwarded
+  to InfluxDB — the stored `peer_id` tag comes from the request body and is constrained
+  only by the tag allowlist and the maximum tag value length.
+- **The InfluxDB token stays server-side** — clients never hold a write credential.
 - **InfluxDB is not exposed** — only accessible within the docker network
 - Source: `ingest/main.go`
 
@@ -56,14 +67,33 @@ Measurement: `netbird_peer_connection`
 
 Tags:
 - `deployment_type`: "cloud" | "selfhosted" | "unknown"
-- `connection_type`: "ice" | "relay"
+- `connection_type`: "ice_p2p" | "ice_turn" | "relay" (see below)
 - `attempt_type`: "initial" | "reconnection"
 - `version`: NetBird version string
 - `os`: Operating system (linux, darwin, windows, android, ios, etc.)
 - `arch`: CPU architecture (amd64, arm64, etc.)
+- `peer_id`: obfuscated peer identifier (truncated SHA-256 of the WireGuard public key)
+- `connection_pair_id`: deterministic identifier for the peer pair, identical on both sides
 
 **Note:** `SignalingReceived` is set when the first offer or answer arrives from the remote peer (in both initial and reconnection paths). It excludes the potentially unbounded wait for the remote peer to come online.
 
+#### `connection_type` values
+
+Derived from the connection priority (`conntype.ConnPriority`) by `metricsConnType` in `client/internal/peer/conn.go`:
+
+| Value | Priority | Traffic is |
+|-------|----------|------------|
+| `ice_p2p` | `ICEP2P` | direct peer-to-peer |
+| `ice_turn` | `ICETurn` | relayed, through a TURN server |
+| `relay` | `Relay` | relayed, through a NetBird relay |
+| `unknown` | `None` or unrecognised | no active transport — **the sample is not pushed** |
+
+**Direct traffic is `ice_p2p` only.** `ice_turn` is relayed despite being negotiated by ICE, matching `Conn.isRelayed`.
+
+`None` means no transport is active: not established yet, or reset after a relay drop or a peer-state reset. Such a sample cannot be attributed to a transport, so `recordConnectionMetrics` drops it instead of pushing it — `unknown` therefore never appears in the bucket. Connection counts are counts of connections whose transport was known at sampling time.
+
+**Samples recorded before 0.77 used a single `ice` value** which covered `ICEP2P`, `ICETurn` *and* `None`, so historical `ice` samples overstate direct connections by an unknown amount and must not be compared with `ice_p2p`.
+
 ### Sync Duration
 
 Measurement: `netbird_sync`
@@ -176,7 +206,7 @@ docker compose up -d
 ```
 
 This starts:
-- **Ingest server** on http://localhost:8087 — accepts client metrics (requires `X-Peer-ID` header, no secret/token auth)
+- **Ingest server** on http://localhost:8087 — accepts client metrics (unauthenticated by design; expects a well-formed `X-Peer-ID` correlation tag)
 - **InfluxDB** — internal only, not exposed to host
 - **Grafana** on http://localhost:3001
 

client/internal/metrics/infra/ingest/main.go

@@ -22,6 +22,11 @@ const (
 	maxDurationSeconds = 86400.0          // reject any duration field > 24 hours
 	peerIDLength       = 16               // truncated SHA-256: 8 bytes = 16 hex chars
 	maxTagValueLength  = 64               // reject tag values longer than this
+	readTimeout        = 30 * time.Second // must fit reading a compressed body up to maxBodySize
+	writeTimeout       = 60 * time.Second // must exceed the upstream client timeout below
+	idleTimeout        = 120 * time.Second
+	readHeaderTimeout  = 10 * time.Second
+	maxHeaderBytes     = 1 << 20 // 1 MB
 )
 
 type measurementSpec struct {
@@ -124,8 +129,17 @@ func main() {
 		fmt.Fprint(w, "ok") //nolint:errcheck
 	})
 
+	srv := &http.Server{
+		Addr:              listenAddr,
+		ReadTimeout:       readTimeout,
+		ReadHeaderTimeout: readHeaderTimeout,
+		WriteTimeout:      writeTimeout,
+		IdleTimeout:       idleTimeout,
+		MaxHeaderBytes:    maxHeaderBytes,
+	}
+
 	log.Printf("ingest server listening on %s, forwarding to %s", listenAddr, influxURL)
-	if err := http.ListenAndServe(listenAddr, nil); err != nil { //nolint:gosec
+	if err := srv.ListenAndServe(); err != nil {
 		log.Fatal(err)
 	}
 }
@@ -137,8 +151,8 @@ func handleIngest(client *http.Client, influxURL, influxToken string) http.Handl
 			return
 		}
 
-		if err := validateAuth(r); err != nil {
-			http.Error(w, err.Error(), http.StatusUnauthorized)
+		if err := validatePeerIDFormat(r); err != nil {
+			http.Error(w, err.Error(), http.StatusBadRequest)
 			return
 		}
 
@@ -187,8 +201,13 @@ func forwardToInflux(w http.ResponseWriter, r *http.Request, client *http.Client
 	io.Copy(w, resp.Body) //nolint:errcheck
 }
 
-// validateAuth checks that the X-Peer-ID header contains a valid hashed peer ID.
-func validateAuth(r *http.Request) error {
+// validatePeerIDFormat checks the shape of the X-Peer-ID header. The header is a
+// correlation tag, not a credential: this endpoint is intentionally
+// unauthenticated so that peers of self-hosted deployments, for which no shared
+// trust anchor exists, can report obfuscated telemetry. The header is not forwarded
+// to InfluxDB, so this check does not bound the stored peer_id tag; it only rejects
+// a malformed header as a bad request rather than an auth failure.
+func validatePeerIDFormat(r *http.Request) error {
 	peerID := r.Header.Get("X-Peer-ID")
 	if peerID == "" {
 		return fmt.Errorf("missing X-Peer-ID header")

client/internal/metrics/infra/ingest/main_test.go

@@ -94,7 +94,7 @@ func TestValidateLineProtocol_RejectsOnBadLine(t *testing.T) {
 	require.Error(t, err)
 }
 
-func TestValidateAuth(t *testing.T) {
+func TestValidatePeerIDFormat(t *testing.T) {
 	tests := []struct {
 		name    string
 		peerID  string
@@ -113,7 +113,7 @@ func TestValidateAuth(t *testing.T) {
 			if tt.peerID != "" {
 				r.Header.Set("X-Peer-ID", tt.peerID)
 			}
-			err := validateAuth(r)
+			err := validatePeerIDFormat(r)
 			if tt.wantErr {
 				require.Error(t, err)
 			} else {

client/internal/metrics/metrics.go

@@ -89,6 +89,21 @@ type ConnectionStageTimestamps struct {
 	WgHandshakeSuccess time.Time
 }
 
+// Durations returns the stage durations in seconds. A duration is zero when
+// either of its timestamps is missing.
+func (c ConnectionStageTimestamps) Durations() (signalingToConnection, connectionToWgHandshake, total float64) {
+	if !c.SignalingReceived.IsZero() && !c.ConnectionReady.IsZero() {
+		signalingToConnection = c.ConnectionReady.Sub(c.SignalingReceived).Seconds()
+	}
+	if !c.ConnectionReady.IsZero() && !c.WgHandshakeSuccess.IsZero() {
+		connectionToWgHandshake = c.WgHandshakeSuccess.Sub(c.ConnectionReady).Seconds()
+	}
+	if !c.SignalingReceived.IsZero() && !c.WgHandshakeSuccess.IsZero() {
+		total = c.WgHandshakeSuccess.Sub(c.SignalingReceived).Seconds()
+	}
+	return signalingToConnection, connectionToWgHandshake, total
+}
+
 // String returns a human-readable representation of the connection stage timestamps
 func (c ConnectionStageTimestamps) String() string {
 	return fmt.Sprintf("ConnectionStageTimestamps{SignalingReceived=%v, ConnectionReady=%v, WgHandshakeSuccess=%v}",
@@ -279,3 +294,11 @@ func (c *ClientMetrics) stopPushLocked() {
 	c.wg.Wait()
 	c.push.Store(nil)
 }
+
+// attemptType returns the metric label for an initial vs reconnection attempt.
+func attemptType(isReconnection bool) string {
+	if isReconnection {
+		return "reconnection"
+	}
+	return "initial"
+}

client/internal/metrics/metrics_default.go

@@ -2,10 +2,24 @@
 
 package metrics
 
+import "github.com/prometheus/client_golang/prometheus"
+
 // NewClientMetrics creates a new ClientMetrics instance
 func NewClientMetrics(agentInfo AgentInfo) *ClientMetrics {
 	return &ClientMetrics{
-		impl:      newInfluxDBMetrics(),
+		impl:      newPrometheusMetrics(newInfluxDBMetrics()),
 		agentInfo: agentInfo,
 	}
 }
+
+// PrometheusGatherer returns the registry with the mirrored Prometheus
+// metrics, or nil when unavailable.
+func (c *ClientMetrics) PrometheusGatherer() prometheus.Gatherer {
+	if c == nil {
+		return nil
+	}
+	if pm, ok := c.impl.(*prometheusMetrics); ok {
+		return pm.Gatherer()
+	}
+	return nil
+}

client/internal/metrics/prometheus.go

@@ -0,0 +1,119 @@
+//go:build !js
+
+package metrics
+
+import (
+	"context"
+	"io"
+	"strconv"
+	"time"
+
+	"github.com/prometheus/client_golang/prometheus"
+)
+
+// prometheusMetrics mirrors recorded client metrics into a Prometheus
+// registry for the local /metrics endpoint, then delegates to the wrapped
+// implementation. Export and Reset pass through untouched: Prometheus
+// metrics are cumulative and pull-based.
+type prometheusMetrics struct {
+	next     metricsImplementation
+	registry *prometheus.Registry
+
+	connectionStages  *prometheus.HistogramVec
+	syncDuration      prometheus.Histogram
+	syncPhaseDuration *prometheus.HistogramVec
+	loginDuration     *prometheus.HistogramVec
+}
+
+func newPrometheusMetrics(next metricsImplementation) *prometheusMetrics {
+	connectionBuckets := []float64{.05, .1, .25, .5, 1, 2.5, 5, 10, 30, 60}
+
+	m := &prometheusMetrics{
+		next:     next,
+		registry: prometheus.NewRegistry(),
+		connectionStages: prometheus.NewHistogramVec(prometheus.HistogramOpts{
+			Name:    "netbird_peer_connection_stage_duration_seconds",
+			Help:    "Duration of peer connection establishment stages.",
+			Buckets: connectionBuckets,
+		}, []string{"stage", "connection_type", "attempt_type"}),
+		syncDuration: prometheus.NewHistogram(prometheus.HistogramOpts{
+			Name:    "netbird_sync_duration_seconds",
+			Help:    "Duration of management sync message processing.",
+			Buckets: prometheus.DefBuckets,
+		}),
+		syncPhaseDuration: prometheus.NewHistogramVec(prometheus.HistogramOpts{
+			Name:    "netbird_sync_phase_duration_seconds",
+			Help:    "Duration of individual sync processing phases.",
+			Buckets: prometheus.DefBuckets,
+		}, []string{"phase"}),
+		loginDuration: prometheus.NewHistogramVec(prometheus.HistogramOpts{
+			Name:    "netbird_login_duration_seconds",
+			Help:    "Duration of logins to the management service.",
+			Buckets: prometheus.DefBuckets,
+		}, []string{"success"}),
+	}
+
+	m.registry.MustRegister(m.connectionStages, m.syncDuration, m.syncPhaseDuration, m.loginDuration)
+	return m
+}
+
+// Gatherer returns the registry holding the mirrored metrics.
+func (m *prometheusMetrics) Gatherer() prometheus.Gatherer {
+	return m.registry
+}
+
+// RecordConnectionStages implements metricsImplementation.
+func (m *prometheusMetrics) RecordConnectionStages(
+	ctx context.Context,
+	agentInfo AgentInfo,
+	connectionPairID string,
+	connectionType ConnectionType,
+	isReconnection bool,
+	timestamps ConnectionStageTimestamps,
+) {
+	attempt := attemptType(isReconnection)
+	connType := connectionType.String()
+
+	signalingToConnection, connectionToWgHandshake, total := timestamps.Durations()
+	if signalingToConnection > 0 {
+		m.connectionStages.WithLabelValues("signaling_to_connection", connType, attempt).Observe(signalingToConnection)
+	}
+	if connectionToWgHandshake > 0 {
+		m.connectionStages.WithLabelValues("connection_to_wg_handshake", connType, attempt).Observe(connectionToWgHandshake)
+	}
+	if total > 0 {
+		m.connectionStages.WithLabelValues("total", connType, attempt).Observe(total)
+	}
+
+	m.next.RecordConnectionStages(ctx, agentInfo, connectionPairID, connectionType, isReconnection, timestamps)
+}
+
+// RecordSyncDuration implements metricsImplementation.
+func (m *prometheusMetrics) RecordSyncDuration(ctx context.Context, agentInfo AgentInfo, duration time.Duration) {
+	m.syncDuration.Observe(duration.Seconds())
+	m.next.RecordSyncDuration(ctx, agentInfo, duration)
+}
+
+// RecordSyncPhase implements metricsImplementation.
+func (m *prometheusMetrics) RecordSyncPhase(ctx context.Context, agentInfo AgentInfo, phase string, duration time.Duration) {
+	m.syncPhaseDuration.WithLabelValues(phase).Observe(duration.Seconds())
+	m.next.RecordSyncPhase(ctx, agentInfo, phase, duration)
+}
+
+// RecordLoginDuration implements metricsImplementation.
+func (m *prometheusMetrics) RecordLoginDuration(ctx context.Context, agentInfo AgentInfo, duration time.Duration, success bool) {
+	m.loginDuration.WithLabelValues(strconv.FormatBool(success)).Observe(duration.Seconds())
+	m.next.RecordLoginDuration(ctx, agentInfo, duration, success)
+}
+
+// Export implements metricsImplementation by delegating to the wrapped
+// implementation; Prometheus metrics are pulled via the registry instead.
+func (m *prometheusMetrics) Export(w io.Writer) error {
+	return m.next.Export(w)
+}
+
+// Reset implements metricsImplementation by delegating to the wrapped
+// implementation; Prometheus metrics must not be cleared on push.
+func (m *prometheusMetrics) Reset() {
+	m.next.Reset()
+}

client/internal/peer/conn.go

@@ -30,6 +30,7 @@ import (
 	"github.com/netbirdio/netbird/client/internal/portforward"
 	"github.com/netbirdio/netbird/client/internal/rosenpass"
 	"github.com/netbirdio/netbird/client/internal/stdnet"
+	"github.com/netbirdio/netbird/client/netevents"
 	"github.com/netbirdio/netbird/route"
 	relayClient "github.com/netbirdio/netbird/shared/relay/client"
 )
@@ -104,6 +105,10 @@ type ConnConfig struct {
 
 	// ICEConfig ICE protocol configuration
 	ICEConfig icemaker.Config
+
+	// NetMgr gates the reconnection guard on OS-reported network
+	// availability; nil disables gating.
+	NetMgr *netevents.Manager
 }
 
 func (c ConnConfig) IsController() bool {
@@ -265,7 +270,7 @@ func (conn *Conn) open(engineCtx context.Context, firstPacket []byte) error {
 		RosenpassAddr:   conn.config.RosenpassConfig.Addr,
 	}, conn.signaler, iceWorker, conn.relayManager)
 
-	conn.guard = guard.NewGuard(conn.Log, conn.isConnectedOnAllWay, conn.config.Timeout, conn.srWatcher)
+	conn.guard = guard.NewGuard(conn.Log, conn.isConnectedOnAllWay, conn.config.Timeout, conn.srWatcher, conn.config.NetMgr)
 
 	conn.relayDialInFlight = false
 	conn.pendingRelayOffer = nil
@@ -481,6 +486,7 @@ func (conn *Conn) teardown(mb *mailbox, leftover []event, signalToRemote bool, d
 
 	if conn.wgWatcherCancel != nil {
 		conn.wgWatcherCancel()
+		conn.wgWatcher = nil
 		conn.wgWatcherCancel = nil
 	}
 	conn.workerRelay.CloseConn()
@@ -650,7 +656,7 @@ func (conn *Conn) handleICEReady(priority worker.ConnPriority, iceConnInfo worke
 		conn.dumpState.NewLocalProxy()
 		wgProxy, err = conn.newProxy(iceConnInfo.RemoteConn)
 		if err != nil {
-			conn.Log.Errorf("failed to add turn net.Conn to local proxy: %v", err)
+			conn.Log.Errorf("failed to add relayed net.Conn to local proxy: %v", err)
 			return
 		}
 		ep = wgProxy.EndpointAddr()
@@ -1142,9 +1148,8 @@ func (conn *Conn) newProxy(remoteConn net.Conn) (wgproxy.Proxy, error) {
 	}
 
 	wgProxy := conn.config.WgConfig.WgInterface.GetProxy()
-	if err := wgProxy.AddTurnConn(conn.ctx, udpAddr, remoteConn); err != nil {
-		conn.Log.Errorf("failed to add turn net.Conn to local proxy: %v", err)
-		return nil, err
+	if err := wgProxy.AddRelayedConn(conn.ctx, udpAddr, remoteConn); err != nil {
+		return nil, fmt.Errorf("add relayed conn to proxy: %w", err)
 	}
 	return wgProxy, nil
 }
@@ -1205,12 +1210,9 @@ func (conn *Conn) recordConnectionMetrics() {
 		return
 	}
 
-	var connType metrics.ConnectionType
-	switch conn.currentConnPriority {
-	case worker.Relay:
-		connType = metrics.ConnectionTypeRelay
-	default:
-		connType = metrics.ConnectionTypeICE
+	connType := metricsConnType(conn.currentConnPriority)
+	if connType == metrics.ConnectionTypeUnknown {
+		return
 	}
 
 	// Record metrics with timestamps - duration calculation happens in metrics package
@@ -1298,3 +1300,16 @@ func boolToConnStatus(connected bool) guard.ConnStatus {
 	}
 	return guard.ConnStatusDisconnected
 }
+
+func metricsConnType(priority worker.ConnPriority) metrics.ConnectionType {
+	switch priority {
+	case worker.Relay:
+		return metrics.ConnectionTypeRelay
+	case worker.ICETurn:
+		return metrics.ConnectionTypeICETurn
+	case worker.ICEP2P:
+		return metrics.ConnectionTypeICEP2P
+	default:
+		return metrics.ConnectionTypeUnknown
+	}
+}

client/internal/peer/conn_signaling_test.go

@@ -0,0 +1,52 @@
+package peer
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/peer/metricsstages"
+	"github.com/netbirdio/netbird/client/internal/peer/signaling"
+)
+
+func TestConn_AnswerBeforeEventLoop(t *testing.T) {
+	for _, tc := range []struct {
+		name  string
+		ports []int
+	}{
+		{name: "holds early answer", ports: []int{51820}},
+		{name: "keeps latest answer", ports: []int{1111, 2222}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			conn, err := NewConn(connConf, ServiceDependencies{})
+			require.NoError(t, err)
+			conn.metricsStages = &metricsstages.MetricsStages{}
+			conn.handshaker = signaling.NewHandshaker(conn.Log, signaling.Config{}, nil, nil, nil)
+			// A relay dial in progress retains the dispatched answer as its next offer.
+			conn.relayDialInFlight = true
+			mb := newMailbox()
+			conn.mailbox.Store(mb)
+
+			// Incoming answers can arrive after Open publishes the mailbox but
+			// before the event loop gets scheduled to consume it.
+			for _, port := range tc.ports {
+				conn.OnRemoteAnswer(signaling.OfferAnswer{WgListenPort: port})
+			}
+
+			select {
+			case <-mb.wake:
+			default:
+				t.Fatal("an early answer must wake the event loop")
+			}
+			events := mb.drain()
+			require.Len(t, events, 1, "only the latest answer should reach the event loop")
+			for _, ev := range events {
+				conn.handleEvent(ev)
+			}
+			require.NotNil(t, conn.pendingRelayOffer, "the answer must reach relay dispatch")
+			assert.Equal(t, tc.ports[len(tc.ports)-1], conn.pendingRelayOffer.WgListenPort,
+				"relay dispatch must receive the latest queued answer")
+		})
+	}
+}

client/internal/peer/conn_test.go

@@ -13,11 +13,13 @@ import (
 	"github.com/stretchr/testify/require"
 
 	"github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/internal/metrics"
 	"github.com/netbirdio/netbird/client/internal/peer/guard"
 	"github.com/netbirdio/netbird/client/internal/peer/ice"
 	"github.com/netbirdio/netbird/client/internal/peer/metricsstages"
 	"github.com/netbirdio/netbird/client/internal/peer/signaling"
 	"github.com/netbirdio/netbird/client/internal/peer/status"
+	"github.com/netbirdio/netbird/client/internal/peer/worker"
 	"github.com/netbirdio/netbird/client/internal/stdnet"
 	"github.com/netbirdio/netbird/util"
 )
@@ -354,3 +356,33 @@ func TestConn_onWGDisconnected_NoEscalationWithoutRosenpass(t *testing.T) {
 	}
 	assert.Empty(t, disconnected, "escalation must be limited to rosenpass connections")
 }
+
+func TestMetricsConnType(t *testing.T) {
+	tests := []struct {
+		name     string
+		priority worker.ConnPriority
+		expected metrics.ConnectionType
+	}{
+		{"relay", worker.Relay, metrics.ConnectionTypeRelay},
+		{"ice over turn is relayed, not p2p", worker.ICETurn, metrics.ConnectionTypeICETurn},
+		{"direct p2p", worker.ICEP2P, metrics.ConnectionTypeICEP2P},
+		{"unset priority is unknown, not p2p", worker.None, metrics.ConnectionTypeUnknown},
+		{"unrecognised priority is unknown", worker.ConnPriority(99), metrics.ConnectionTypeUnknown},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			assert.Equal(t, tc.expected, metricsConnType(tc.priority))
+		})
+	}
+}
+
+func TestMetricsConnType_RelayedMatchesIsRelayed(t *testing.T) {
+	for _, priority := range []worker.ConnPriority{worker.None, worker.Relay, worker.ICETurn, worker.ICEP2P} {
+		conn := &Conn{currentConnPriority: priority}
+		tag := metricsConnType(priority)
+		relayedTag := tag == metrics.ConnectionTypeRelay || tag == metrics.ConnectionTypeICETurn
+		assert.Equal(t, conn.isRelayed(), relayedTag,
+			"priority %s: isRelayed and the %q metric tag must agree", priority, tag)
+	}
+}