Run a single live Kimi scenario: kimi-k3 over the Anthropic shape

The Moonshot platform serves this account exactly one model — kimi-k3.
kimi-k2-thinking and the kimi-latest alias both return
resource_not_found_error on either API surface, so two concurrent Kimi
provider records would claim the same model and route ambiguously
(the router's path-prefix tiebreak needs the client to author the
/anthropic prefix, which the standard wire shapes don't). Keep the
flagship Claude Code path — kimi-k3 via /anthropic, green in run #74 —
as the live scenario; the OpenAI wire shape stays covered by the other
chat-shaped matrix providers and passed with kimi-k3 in run #73.

.github/workflows/agent-network-e2e.yml

@@ -5,6 +5,10 @@ on:
   schedule:
     - cron: "0 3 * * *"
   workflow_dispatch:
+  # TEMPORARY: run on pushes to the Kimi feature branch to exercise the new
+  # provider scenarios with E2E_KIMI_TOKEN. Remove before merging.
+  push:
+    branches: [feat/kimi-3-agent-networks]
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
@@ -51,6 +55,9 @@ jobs:
           # token (and URL, for gateways) is unset, so partial coverage is fine.
           OPENAI_TOKEN: ${{ secrets.E2E_OPENAI_TOKEN }}
           ANTHROPIC_TOKEN: ${{ secrets.E2E_ANTHROPIC_TOKEN }}
+          # Moonshot AI platform key (platform.kimi.ai); drives both Kimi wire
+          # shapes (OpenAI /v1 and Anthropic /anthropic) through kimi_api.
+          KIMI_TOKEN: ${{ secrets.E2E_KIMI_TOKEN }}
           VERCEL_URL: ${{ secrets.E2E_VERCEL_URL }}
           VERCEL_TOKEN: ${{ secrets.E2E_VERCEL_TOKEN }}
           OPENROUTER_URL: ${{ secrets.E2E_OPENROUTER_URL }}

CLAUDE.md

@@ -0,0 +1,6 @@
+# Repository guidance for Claude
+
+## Git conventions
+
+- Do not add `Co-Authored-By` or other AI attribution trailers to commit messages.
+- Do not use `claude/` (or otherwise Claude-named) branch names; use conventional prefixes like `feat/`, `fix/`, `chore/`.

client/android/client.go

@@ -247,6 +247,9 @@ func (c *Client) DebugBundle(platformFiles PlatformFiles, anonymize bool) (strin
 		deps.SyncResponse = resp
 
 		if e := cc.Engine(); e != nil {
+			deps.RefreshStatus = func() {
+				e.RunHealthProbes(context.Background(), true)
+			}
 			if cm := e.GetClientMetrics(); cm != nil {
 				deps.ClientMetrics = cm
 			}

client/cmd/login.go

@@ -17,7 +17,9 @@ import (
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/auth"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
+	nbnet "github.com/netbirdio/netbird/client/net"
 	"github.com/netbirdio/netbird/client/proto"
+	"github.com/netbirdio/netbird/client/server"
 	"github.com/netbirdio/netbird/client/system"
 	"github.com/netbirdio/netbird/util"
 )
@@ -331,6 +333,14 @@ func doForegroundLogin(ctx context.Context, cmd *cobra.Command, setupKey string,
 		return fmt.Errorf("read config file %s: %v", configFilePath, err)
 	}
 
+	// Mirror runInForegroundMode: recover residual state (DNS, firewall,
+	// ssh config, legacy routing) from a previous unclean shutdown and
+	// enable advanced routing before dialing management.
+	if err := server.RestoreResidualState(ctx, profilemanager.NewServiceManager(configFilePath).GetStatePath()); err != nil {
+		log.Warnf("failed to restore residual state: %v", err)
+	}
+	nbnet.Init()
+
 	err = foregroundLogin(ctx, cmd, config, setupKey, activeProf.ID)
 	if err != nil {
 		return fmt.Errorf("foreground login failed: %v", err)

client/cmd/up.go

@@ -22,6 +22,8 @@ import (
 	"github.com/netbirdio/netbird/client/internal/peer"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/proto"
+	nbnet "github.com/netbirdio/netbird/client/net"
+	"github.com/netbirdio/netbird/client/server"
 	"github.com/netbirdio/netbird/client/system"
 	"github.com/netbirdio/netbird/shared/management/domain"
 	"github.com/netbirdio/netbird/util"
@@ -229,6 +231,24 @@ func runInForegroundMode(ctx context.Context, cmd *cobra.Command, activeProf *pr
 
 	_, _ = profilemanager.UpdateOldManagementURL(ctx, config, configFilePath)
 
+	// Restore residual state left by a previous run that did not shut down
+	// cleanly, mirroring what the daemon does before connecting: it recovers
+	// DNS config (a stale resolv.conf takeover can make the management
+	// hostname unresolvable), firewall rules, ssh config and legacy routing.
+	// Route cleanup itself happens at engine start; nbnet.Init() below lets
+	// the management dial bypass a leftover fwmark rule until then.
+	// Foreground mode is particularly exposed in containers: a crashed
+	// container restarts inside the same (pod) network namespace, so stale
+	// state survives while the process does not.
+	if err := server.RestoreResidualState(ctx, profilemanager.NewServiceManager(configPath).GetStatePath()); err != nil {
+		log.Warnf("failed to restore residual state: %v", err)
+	}
+
+	// Enable advanced routing (as the daemon does on startup) so the
+	// management dial bypasses a leftover fwmark rule instead of being
+	// shunted into a stale routing table.
+	nbnet.Init()
+
 	err = foregroundLogin(ctx, cmd, config, providedSetupKey, activeProf.ID)
 	if err != nil {
 		return fmt.Errorf("foreground login failed: %v", err)

client/firewall/uspfilter/filter.go

@@ -121,6 +121,7 @@ type Manager struct {
 	udpTracker     *conntrack.UDPTracker
 	icmpTracker    *conntrack.ICMPTracker
 	tcpTracker     *conntrack.TCPTracker
+	fragments      *fragmentTracker
 	forwarder      atomic.Pointer[forwarder.Forwarder]
 	pendingCapture atomic.Pointer[forwarder.PacketCapture]
 	logger         *nblog.Logger
@@ -183,6 +184,41 @@ func (d *decoder) decodePacket(data []byte) error {
 	}
 }
 
+// decodeTransport decodes the transport header of a first fragment (which
+// gopacket leaves undecoded) into the decoder and appends its layer type to
+// decoded, so the ACL pipeline can evaluate it like a normal packet. It returns
+// false if the protocol is unsupported or the header is truncated.
+func (d *decoder) decodeTransport(proto layers.IPProtocol, payload []byte) bool {
+	var l4 gopacket.DecodingLayer
+	var layerType gopacket.LayerType
+	var minLen int
+	switch proto {
+	case layers.IPProtocolTCP:
+		l4, layerType, minLen = &d.tcp, layers.LayerTypeTCP, 20
+	case layers.IPProtocolUDP:
+		l4, layerType, minLen = &d.udp, layers.LayerTypeUDP, 8
+	case layers.IPProtocolICMPv4:
+		l4, layerType, minLen = &d.icmp4, layers.LayerTypeICMPv4, 8
+	case layers.IPProtocolICMPv6:
+		l4, layerType, minLen = &d.icmp6, layers.LayerTypeICMPv6, 8
+	default:
+		return false
+	}
+
+	// Reject a fragment too small to hold the full transport header before
+	// decoding: it can't be ACL-evaluated (tiny-fragment attack), and skipping
+	// the decode avoids gopacket allocating an error on the drop path.
+	if len(payload) < minLen {
+		return false
+	}
+
+	if err := l4.DecodeFromBytes(payload, gopacket.NilDecodeFeedback); err != nil {
+		return false
+	}
+	d.decoded = append(d.decoded, layerType)
+	return true
+}
+
 // Create userspace firewall manager constructor
 func Create(iface common.IFaceMapper, disableServerRoutes bool, flowLogger nftypes.FlowLogger, mtu uint16) (*Manager, error) {
 	return create(iface, nil, disableServerRoutes, flowLogger, mtu)
@@ -286,6 +322,8 @@ func create(iface common.IFaceMapper, nativeFirewall firewall.Manager, disableSe
 	if err := m.localipmanager.UpdateLocalIPs(iface); err != nil {
 		return nil, fmt.Errorf("update local IPs: %w", err)
 	}
+	m.fragments = newFragmentTracker(m.logger)
+
 	if disableConntrack {
 		log.Info("conntrack is disabled")
 	} else {
@@ -299,6 +337,7 @@ func create(iface common.IFaceMapper, nativeFirewall firewall.Manager, disableSe
 		}
 	}
 	if err := iface.SetFilter(m); err != nil {
+		m.fragments.Close()
 		return nil, fmt.Errorf("set filter: %w", err)
 	}
 	return m, nil
@@ -694,6 +733,10 @@ func (m *Manager) resetState() {
 		m.tcpTracker.Close()
 	}
 
+	if m.fragments != nil {
+		m.fragments.Close()
+	}
+
 	if fwder := m.forwarder.Load(); fwder != nil {
 		fwder.SetCapture(nil)
 		fwder.Stop()
@@ -1046,19 +1089,20 @@ func (m *Manager) filterInbound(packetData []byte, size int) bool {
 		return true
 	}
 
-	// TODO: pass fragments of routed packets to forwarder
+	// gopacket does not decode the transport header of any IP fragment, so
+	// fragments take a dedicated path: the first fragment's header is decoded
+	// and ACL-evaluated here, and the remaining fragments inherit its verdict.
 	if fragment {
-		if m.logger.Enabled(nblog.LevelTrace) {
-			if d.decoded[0] == layers.LayerTypeIPv4 {
-				m.logger.Trace4("packet is a fragment: src=%v dst=%v id=%v flags=%v",
-					srcIP, dstIP, d.ip4.Id, d.ip4.Flags)
-			} else {
-				m.logger.Trace2("packet is an IPv6 fragment: src=%v dst=%v", srcIP, dstIP)
-			}
-		}
-		return false
+		return m.filterInboundFragment(d, srcIP, dstIP, size)
 	}
 
+	return m.filterInboundDecoded(d, srcIP, dstIP, packetData, size)
+}
+
+// filterInboundDecoded runs the ACL, DNAT and conntrack pipeline on a fully
+// decoded (non-fragment) inbound packet. It returns true if the packet should
+// be dropped.
+func (m *Manager) filterInboundDecoded(d *decoder, srcIP, dstIP netip.Addr, packetData []byte, size int) bool {
 	// TODO: optimize port DNAT by caching matched rules in conntrack
 	if translated := m.translateInboundPortDNAT(packetData, d, srcIP, dstIP); translated {
 		// Re-decode after port DNAT translation to update port information
@@ -1089,33 +1133,226 @@ func (m *Manager) filterInbound(packetData []byte, size int) bool {
 	return m.handleRoutedTraffic(d, srcIP, dstIP, packetData, size)
 }
 
+// fragmentMeta holds the reassembly identity and layout of an IP fragment,
+// extracted uniformly for IPv4 and IPv6.
+type fragmentMeta struct {
+	key fragmentKey
+	// offset is the fragment offset in 8-byte units (zero for the first
+	// fragment).
+	offset uint16
+	// moreFragments is the More Fragments bit. A first fragment with it unset is
+	// an IPv6 atomic fragment (a complete datagram, RFC 6946): it has no trailing
+	// fragments to inherit a verdict, so it must not be recorded.
+	moreFragments bool
+	proto         layers.IPProtocol
+	// l4payload is the fragmentable payload of this fragment. For the first
+	// fragment it starts with the transport header.
+	l4payload []byte
+	// headerEndOctets is the first fragment's payload length in 8-byte units:
+	// the smallest offset a trailing fragment may start at without overlapping
+	// the inspected transport header.
+	headerEndOctets uint16
+}
+
+// fragmentMetadata extracts the fragment identity and layout from a decoded IP
+// fragment. It returns false for fragments it can't interpret (e.g. an IPv6
+// fragment header shorter than 8 bytes), which are then dropped.
+func fragmentMetadata(d *decoder, srcIP, dstIP netip.Addr) (fragmentMeta, bool) {
+	switch d.decoded[0] {
+	case layers.LayerTypeIPv4:
+		payload := d.ip4.Payload
+		return fragmentMeta{
+			key:             fragmentKey{srcIP: srcIP, dstIP: dstIP, id: uint32(d.ip4.Id), proto: uint8(d.ip4.Protocol)},
+			offset:          d.ip4.FragOffset,
+			moreFragments:   d.ip4.Flags&layers.IPv4MoreFragments != 0,
+			proto:           d.ip4.Protocol,
+			l4payload:       payload,
+			headerEndOctets: octets(len(payload)),
+		}, true
+
+	case layers.LayerTypeIPv6:
+		// IPv6 fragment extension header: 8 bytes, followed by the fragmentable
+		// payload. Layout: next header (1), reserved (1), offset+flags (2), id (4).
+		payload := d.ip6.Payload
+		if len(payload) < 8 {
+			return fragmentMeta{}, false
+		}
+		nextHeader := layers.IPProtocol(payload[0])
+		offsetFlags := binary.BigEndian.Uint16(payload[2:4])
+		id := binary.BigEndian.Uint32(payload[4:8])
+		l4 := payload[8:]
+		return fragmentMeta{
+			key:             fragmentKey{srcIP: srcIP, dstIP: dstIP, id: id, proto: uint8(nextHeader)},
+			offset:          offsetFlags >> 3,
+			moreFragments:   offsetFlags&1 != 0,
+			proto:           nextHeader,
+			l4payload:       l4,
+			headerEndOctets: octets(len(l4)),
+		}, true
+
+	default:
+		return fragmentMeta{}, false
+	}
+}
+
+// octets rounds a byte length up to whole 8-byte units, the granularity of the
+// IP fragment offset field.
+func octets(nbytes int) uint16 {
+	return uint16((nbytes + 7) / 8)
+}
+
+// filterInboundFragment decides the fate of an IP fragment. gopacket stops
+// decoding at the network layer for every fragment, so the first fragment's
+// transport header is decoded and ACL-evaluated here and its verdict recorded;
+// the remaining (headerless) fragments inherit that verdict. Anything that
+// cannot be tied to an allowed, non-overlapping first fragment is dropped.
+func (m *Manager) filterInboundFragment(d *decoder, srcIP, dstIP netip.Addr, size int) bool {
+	meta, ok := fragmentMetadata(d, srcIP, dstIP)
+	if !ok {
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace2("dropping unsupported fragment: src=%v dst=%v", srcIP, dstIP)
+		}
+		return true
+	}
+
+	if meta.offset != 0 {
+		return m.filterTrailingFragment(meta, srcIP, dstIP)
+	}
+
+	// A new first fragment supersedes any recorded verdict for this datagram, so
+	// a re-sent or overlapping offset-zero fragment can't inherit the old one.
+	m.fragments.poison(meta.key)
+
+	// First fragment: decode its transport header so the ACL can evaluate it. A
+	// decode failure means the fragment is too small to hold the full transport
+	// header (RFC 1858 §3 tiny-fragment attack); it can't be evaluated, so drop it.
+	if !d.decodeTransport(meta.proto, meta.l4payload) {
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace3("dropping first fragment without full L4 header: src=%v dst=%v id=%v",
+				srcIP, dstIP, meta.key.id)
+		}
+		return true
+	}
+
+	return m.filterFirstFragment(d, meta, srcIP, dstIP, size)
+}
+
+// filterTrailingFragment applies a recorded first-fragment verdict to a
+// non-first fragment.
+func (m *Manager) filterTrailingFragment(meta fragmentMeta, srcIP, dstIP netip.Addr) bool {
+	switch m.fragments.verdict(meta.key, meta.offset) {
+	case fragmentAllow:
+		return false
+	case fragmentOverlap:
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace3("dropping overlapping fragment rewriting inspected header: src=%v dst=%v id=%v",
+				srcIP, dstIP, meta.key.id)
+		}
+		return true
+	default:
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace3("dropping fragment with no allowed first fragment: src=%v dst=%v id=%v",
+				srcIP, dstIP, meta.key.id)
+		}
+		return true
+	}
+}
+
+// filterFirstFragment runs the verdict part of the inbound pipeline on a first
+// fragment with its transport header decoded. It mirrors filterInboundDecoded
+// but skips DNAT (port rewriting on fragments is unsupported) and forwarder
+// injection (fragments are left to the stack to reassemble, not forwarded).
+// Allowed fragments have their verdict recorded so the datagram's trailing
+// fragments inherit it.
+func (m *Manager) filterFirstFragment(d *decoder, meta fragmentMeta, srcIP, dstIP netip.Addr, size int) bool {
+	if m.stateful && m.isValidTrackedConnection(d, srcIP, dstIP, size) {
+		m.recordFirstFragment(meta)
+		return false
+	}
+
+	if m.localipmanager.IsLocalIP(dstIP) {
+		ruleID, blocked := m.peerACLsBlock(srcIP, d, nil)
+		if blocked {
+			m.storeDropFlow("Dropping local first fragment (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
+				d, srcIP, dstIP, ruleID, size)
+			return true
+		}
+		m.trackInbound(d, srcIP, dstIP, ruleID, size)
+		m.recordFirstFragment(meta)
+		return false
+	}
+
+	if !m.routingEnabled.Load() {
+		if m.logger.Enabled(nblog.LevelTrace) {
+			m.logger.Trace2("Dropping routed fragment (routing disabled): src=%s dst=%s", srcIP, dstIP)
+		}
+		return true
+	}
+	if m.nativeRouter.Load() {
+		m.trackInbound(d, srcIP, dstIP, nil, size)
+		m.recordFirstFragment(meta)
+		return false
+	}
+
+	// TODO: pass fragments of routed packets to the forwarder; until then
+	// allowed routed fragments go to the native stack.
+	srcPort, dstPort := getPortsFromPacket(d)
+	ruleID, pass := m.routeACLsPass(srcIP, dstIP, d.decoded[1], srcPort, dstPort)
+	if !pass {
+		m.storeDropFlow("Dropping routed first fragment (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
+			d, srcIP, dstIP, ruleID, size)
+		return true
+	}
+
+	m.recordFirstFragment(meta)
+	return false
+}
+
+// recordFirstFragment caches an allowed first fragment's verdict for its
+// trailing fragments to inherit. Atomic fragments (no More Fragments bit) are
+// complete datagrams with no trailing fragments, so they are not cached and
+// cannot exhaust the verdict table.
+func (m *Manager) recordFirstFragment(meta fragmentMeta) {
+	if !meta.moreFragments {
+		return
+	}
+	m.fragments.recordAllowed(meta.key, meta.headerEndOctets)
+}
+
+// storeDropFlow logs and records a netflow drop event for an inbound packet
+// denied by the ACLs. msg is the trace format taking rule id, protocol, source
+// and destination.
+func (m *Manager) storeDropFlow(msg string, d *decoder, srcIP, dstIP netip.Addr, ruleID []byte, size int) {
+	pnum := getProtocolFromPacket(d)
+	srcPort, dstPort := getPortsFromPacket(d)
+
+	if m.logger.Enabled(nblog.LevelTrace) {
+		m.logger.Trace6(msg, ruleID, pnum, srcIP, srcPort, dstIP, dstPort)
+	}
+
+	m.flowLogger.StoreEvent(nftypes.EventFields{
+		FlowID:     uuid.New(),
+		Type:       nftypes.TypeDrop,
+		RuleID:     ruleID,
+		Direction:  nftypes.Ingress,
+		Protocol:   pnum,
+		SourceIP:   srcIP,
+		DestIP:     dstIP,
+		SourcePort: srcPort,
+		DestPort:   dstPort,
+		// TODO: icmp type/code
+		RxPackets: 1,
+		RxBytes:   uint64(size),
+	})
+}
+
 // handleLocalTraffic handles local traffic.
 // If it returns true, the packet should be dropped.
 func (m *Manager) handleLocalTraffic(d *decoder, srcIP, dstIP netip.Addr, packetData []byte, size int) bool {
 	ruleID, blocked := m.peerACLsBlock(srcIP, d, packetData)
 	if blocked {
-		pnum := getProtocolFromPacket(d)
-		srcPort, dstPort := getPortsFromPacket(d)
-
-		if m.logger.Enabled(nblog.LevelTrace) {
-			m.logger.Trace6("Dropping local packet (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
-				ruleID, pnum, srcIP, srcPort, dstIP, dstPort)
-		}
-
-		m.flowLogger.StoreEvent(nftypes.EventFields{
-			FlowID:     uuid.New(),
-			Type:       nftypes.TypeDrop,
-			RuleID:     ruleID,
-			Direction:  nftypes.Ingress,
-			Protocol:   pnum,
-			SourceIP:   srcIP,
-			DestIP:     dstIP,
-			SourcePort: srcPort,
-			DestPort:   dstPort,
-			// TODO: icmp type/code
-			RxPackets: 1,
-			RxBytes:   uint64(size),
-		})
+		m.storeDropFlow("Dropping local packet (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
+			d, srcIP, dstIP, ruleID, size)
 		return true
 	}
 
@@ -1168,27 +1405,8 @@ func (m *Manager) handleRoutedTraffic(d *decoder, srcIP, dstIP netip.Addr, packe
 
 	ruleID, pass := m.routeACLsPass(srcIP, dstIP, protoLayer, srcPort, dstPort)
 	if !pass {
-		proto := getProtocolFromPacket(d)
-
-		if m.logger.Enabled(nblog.LevelTrace) {
-			m.logger.Trace6("Dropping routed packet (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
-				ruleID, proto, srcIP, srcPort, dstIP, dstPort)
-		}
-
-		m.flowLogger.StoreEvent(nftypes.EventFields{
-			FlowID:     uuid.New(),
-			Type:       nftypes.TypeDrop,
-			RuleID:     ruleID,
-			Direction:  nftypes.Ingress,
-			Protocol:   proto,
-			SourceIP:   srcIP,
-			DestIP:     dstIP,
-			SourcePort: srcPort,
-			DestPort:   dstPort,
-			// TODO: icmp type/code
-			RxPackets: 1,
-			RxBytes:   uint64(size),
-		})
+		m.storeDropFlow("Dropping routed packet (ACL denied): rule_id=%s proto=%v src=%s:%d dst=%s:%d",
+			d, srcIP, dstIP, ruleID, size)
 		return true
 	}
 

client/firewall/uspfilter/forwarder/forwarder.go

@@ -5,7 +5,9 @@ import (
 	"fmt"
 	"net"
 	"net/netip"
+	"os"
 	"runtime"
+	"strconv"
 	"sync"
 	"time"
 
@@ -31,6 +33,11 @@ const (
 	defaultMaxInFlight   = 1024
 	iosReceiveWindow     = 16384
 	iosMaxInFlight       = 256
+
+	// envForceTCPRACK overrides the platform default for gVisor's RACK loss
+	// detection. Set to a truthy value to force RACK on, or a falsy value to
+	// force it off, on any platform.
+	envForceTCPRACK = "NB_FORCE_TCP_RACK"
 )
 
 type Forwarder struct {
@@ -152,6 +159,8 @@ func New(iface common.IFaceMapper, logger *nblog.Logger, flowLogger nftypes.Flow
 		maxInFlight = iosMaxInFlight
 	}
 
+	configureTCPRecovery(s)
+
 	tcpForwarder := tcp.NewForwarder(s, receiveWindow, maxInFlight, f.handleTCP)
 	s.SetTransportProtocolHandler(tcp.ProtocolNumber, tcpForwarder.HandlePacket)
 
@@ -466,3 +475,31 @@ func probeRawICMP(network, addr string, logger *nblog.Logger) bool {
 	logger.Debug1("forwarder: raw %s socket access available", network)
 	return true
 }
+
+// configureTCPRecovery disables gVisor's RACK loss detection on Windows, where
+// it interacts poorly with the host and collapses throughput on routed TCP
+// connections (gVisor issue #9778). Other platforms keep the default. The
+// EnvForceTCPRACK environment variable overrides the platform default.
+func configureTCPRecovery(s *stack.Stack) {
+	disableRACK := runtime.GOOS == "windows"
+
+	if val := os.Getenv(envForceTCPRACK); val != "" {
+		force, err := strconv.ParseBool(val)
+		if err != nil {
+			log.Warnf("parse %s: %v", envForceTCPRACK, err)
+		} else {
+			disableRACK = !force
+		}
+	}
+
+	if !disableRACK {
+		return
+	}
+
+	opt := tcpip.TCPRecovery(0)
+	if err := s.SetTransportProtocolOption(tcp.ProtocolNumber, &opt); err != nil {
+		log.Warnf("disable TCP RACK loss detection: %v", err)
+		return
+	}
+	log.Info("forwarder: TCP RACK loss detection disabled")
+}

client/firewall/uspfilter/fragment.go

@@ -0,0 +1,204 @@
+package uspfilter
+
+import (
+	"context"
+	"net/netip"
+	"os"
+	"strconv"
+	"sync"
+	"time"
+
+	nblog "github.com/netbirdio/netbird/client/firewall/uspfilter/log"
+)
+
+const (
+	// defaultFragmentTimeout bounds how long a first-fragment verdict is kept
+	// while the remaining fragments arrive. It mirrors the Linux IP reassembly
+	// timeout (net.ipv4.ipfrag_time).
+	defaultFragmentTimeout = 30 * time.Second
+	// fragmentCleanupInterval is how often expired verdicts are purged.
+	fragmentCleanupInterval = 10 * time.Second
+	// defaultMaxFragmentEntries caps the number of concurrently tracked
+	// fragmented datagrams. The table stays bounded because each datagram is a
+	// single small entry regardless of how many fragments it is split into, and
+	// the 13-bit IPv4 fragment-offset field limits any datagram to 64 KiB.
+	defaultMaxFragmentEntries = 16384
+
+	// EnvFragmentMaxEntries overrides defaultMaxFragmentEntries.
+	EnvFragmentMaxEntries = "NB_FRAGMENT_MAX_ENTRIES"
+)
+
+// fragmentVerdict is the decision for a trailing (headerless) fragment.
+type fragmentVerdict int
+
+const (
+	// fragmentDeny drops the fragment: no allowed first fragment is on record.
+	fragmentDeny fragmentVerdict = iota
+	// fragmentAllow passes the fragment: it belongs to an allowed datagram and
+	// does not overlap the already-inspected transport header.
+	fragmentAllow
+	// fragmentOverlap drops the fragment and poisons its datagram: it overlaps
+	// the transport header the ACL inspected (RFC 1858 §4, RFC 3128; RFC 5722
+	// requires discarding the whole datagram on overlap for IPv6).
+	fragmentOverlap
+)
+
+// fragmentKey identifies a fragmented datagram. It matches the RFC 791 / RFC
+// 8200 reassembly key: source, destination, protocol and identification. The id
+// is 32-bit to hold both the IPv4 (16-bit) and IPv6 (32-bit) identification.
+type fragmentKey struct {
+	srcIP netip.Addr
+	dstIP netip.Addr
+	id    uint32
+	proto uint8
+}
+
+// fragmentEntry records the verdict of an allowed first fragment.
+type fragmentEntry struct {
+	// headerEndOctets is the offset, in 8-byte units, at which the first
+	// fragment's payload ended. A trailing fragment starting before this
+	// overlaps bytes the ACL already inspected and is rejected.
+	headerEndOctets uint16
+	// recordedAt is when the first fragment was accepted. The verdict expires a
+	// fixed timeout later and is not refreshed, mirroring the kernel reassembly
+	// timer so a trailing-fragment flood can't keep a datagram alive.
+	recordedAt time.Time
+}
+
+// fragmentTracker records the ACL verdict of a datagram's first fragment so the
+// remaining fragments, which carry no L4 header, can inherit the decision
+// without reassembling the datagram. Only allowed first fragments are stored;
+// anything that cannot be tied to an allowed, non-overlapping first fragment is
+// dropped (fail closed).
+type fragmentTracker struct {
+	logger  *nblog.Logger
+	mutex   sync.Mutex
+	entries map[fragmentKey]fragmentEntry
+	timeout time.Duration
+	// maxEntries caps the table; atCapacity dedups the capacity warning until
+	// the table drains below the cap again.
+	maxEntries    int
+	atCapacity    bool
+	cleanupTicker *time.Ticker
+	cancel        context.CancelFunc
+}
+
+func newFragmentTracker(logger *nblog.Logger) *fragmentTracker {
+	ctx, cancel := context.WithCancel(context.Background())
+	t := &fragmentTracker{
+		logger:        logger,
+		entries:       make(map[fragmentKey]fragmentEntry),
+		timeout:       defaultFragmentTimeout,
+		maxEntries:    fragmentMaxEntries(logger),
+		cleanupTicker: time.NewTicker(fragmentCleanupInterval),
+		cancel:        cancel,
+	}
+	go t.cleanupRoutine(ctx)
+	return t
+}
+
+func fragmentMaxEntries(logger *nblog.Logger) int {
+	v := os.Getenv(EnvFragmentMaxEntries)
+	if v == "" {
+		return defaultMaxFragmentEntries
+	}
+	n, err := strconv.Atoi(v)
+	if err != nil || n <= 0 {
+		logger.Warn2("invalid %s=%q, using default", EnvFragmentMaxEntries, v)
+		return defaultMaxFragmentEntries
+	}
+	return n
+}
+
+// recordAllowed stores the verdict of an allowed first fragment. headerEndOctets
+// is the first fragment's payload length in 8-byte units. When the table is full
+// the record is dropped, which fails closed: the datagram's trailing fragments
+// will be denied.
+func (t *fragmentTracker) recordAllowed(key fragmentKey, headerEndOctets uint16) {
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+
+	if t.entries == nil {
+		return
+	}
+	if _, ok := t.entries[key]; !ok && len(t.entries) >= t.maxEntries {
+		if !t.atCapacity {
+			t.atCapacity = true
+			t.logger.Warn2("fragment verdict table at capacity (%d/%d): trailing fragments of new datagrams will be dropped",
+				len(t.entries), t.maxEntries)
+		}
+		return
+	}
+	t.entries[key] = fragmentEntry{
+		headerEndOctets: headerEndOctets,
+		recordedAt:      time.Now(),
+	}
+}
+
+// poison drops any recorded verdict for a datagram, so its later fragments are
+// denied until a new allowed first fragment is recorded. Called on every
+// offset-zero fragment to defeat offset-zero overlap rewrites (RFC 3128).
+func (t *fragmentTracker) poison(key fragmentKey) {
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+	delete(t.entries, key)
+}
+
+// verdict decides the fate of a trailing fragment at fragOffsetOctets (the IPv4
+// fragment offset, in 8-byte units). A fragment overlapping the inspected
+// header poisons the datagram: the entry is removed so all further fragments of
+// that datagram are denied too.
+func (t *fragmentTracker) verdict(key fragmentKey, fragOffsetOctets uint16) fragmentVerdict {
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+
+	entry, ok := t.entries[key]
+	if !ok {
+		return fragmentDeny
+	}
+	if time.Since(entry.recordedAt) > t.timeout {
+		delete(t.entries, key)
+		return fragmentDeny
+	}
+	if fragOffsetOctets < entry.headerEndOctets {
+		delete(t.entries, key)
+		return fragmentOverlap
+	}
+	return fragmentAllow
+}
+
+func (t *fragmentTracker) cleanupRoutine(ctx context.Context) {
+	defer t.cleanupTicker.Stop()
+	for {
+		select {
+		case <-t.cleanupTicker.C:
+			t.cleanup()
+		case <-ctx.Done():
+			return
+		}
+	}
+}
+
+func (t *fragmentTracker) cleanup() {
+	t.mutex.Lock()
+	defer t.mutex.Unlock()
+
+	for key, entry := range t.entries {
+		if time.Since(entry.recordedAt) > t.timeout {
+			delete(t.entries, key)
+		}
+	}
+
+	if len(t.entries) < t.maxEntries {
+		t.atCapacity = false
+	}
+}
+
+// Close stops the cleanup routine and releases resources.
+func (t *fragmentTracker) Close() {
+	t.cancel()
+
+	t.mutex.Lock()
+	t.entries = nil
+	t.mutex.Unlock()
+}

client/firewall/uspfilter/fragment_bench_test.go

@@ -0,0 +1,115 @@
+package uspfilter
+
+import (
+	"encoding/binary"
+	"testing"
+)
+
+// benchFilterInbound drives filterInbound over a fixed packet in a tight loop.
+// Packets are built once, outside the timed region, so the benchmark measures
+// only pipeline cost, which is what an attacker can amplify.
+func benchFilterInbound(b *testing.B, pkt []byte) {
+	b.Helper()
+	b.ReportAllocs()
+	b.SetBytes(int64(len(pkt)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		m := benchManager
+		m.filterInbound(pkt, len(pkt))
+	}
+}
+
+// benchManager is a package-level manager reused across fragment benchmarks so
+// setup cost stays out of the timed region.
+var benchManager *Manager
+
+func setupBenchManager(b *testing.B) *Manager {
+	b.Helper()
+	m := newFragmentTestManager(b)
+	allowUDP(b, m, 8080)
+	// Disable conntrack so the allowed-first-fragment path measures transport
+	// decode + ACL every iteration instead of matching the connection tracked
+	// on the first iteration.
+	m.stateful = false
+	benchManager = m
+	return m
+}
+
+// BenchmarkInbound_NormalPacket is the baseline: a full, non-fragmented UDP
+// packet that passes the ACL. Fragment paths should stay comparable to this.
+func BenchmarkInbound_NormalPacket(b *testing.B) {
+	setupBenchManager(b)
+	pkt := normalUDPPacket(b, 8080, 32)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_FirstFragmentAllowed measures the first-fragment path:
+// transport decode + ACL evaluation + verdict record.
+func BenchmarkInbound_FirstFragmentAllowed(b *testing.B) {
+	setupBenchManager(b)
+	pkt := firstFragmentUDP(b, 0x2000, 8080, 32)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_TrailingFragmentAllowed measures the common trailing-fragment
+// path: a single map lookup after the first fragment is on record.
+func BenchmarkInbound_TrailingFragmentAllowed(b *testing.B) {
+	m := setupBenchManager(b)
+	first := firstFragmentUDP(b, 0x3000, 8080, 32)
+	m.filterInbound(first, len(first))
+	pkt := trailingFragment(b, 0x3000, 5, false, 24)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_TrailingFragmentNoFirst is the primary DoS vector: an
+// attacker floods trailing fragments with no first fragment on record. Each is
+// a map miss and must be cheap.
+func BenchmarkInbound_TrailingFragmentNoFirst(b *testing.B) {
+	setupBenchManager(b)
+	pkt := trailingFragment(b, 0x4000, 185, false, 40)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_TinyFirstFragment measures the tiny-fragment drop path: a
+// first fragment too small to decode a transport header.
+func BenchmarkInbound_TinyFirstFragment(b *testing.B) {
+	setupBenchManager(b)
+	pkt := trailingFragment(b, 0x5000, 0, true, 4)
+	benchFilterInbound(b, pkt)
+}
+
+// BenchmarkInbound_TrailingFragmentDistinctIDs is the worst case for the
+// verdict table: an attacker varies the datagram id on every packet so no first
+// fragment ever matches. Verdict lookups always miss and nothing is recorded,
+// so the table cannot grow. Each iteration rewrites the id field in place.
+func BenchmarkInbound_TrailingFragmentDistinctIDs(b *testing.B) {
+	setupBenchManager(b)
+	pkt := trailingFragment(b, 0x6000, 185, false, 40)
+	m := benchManager
+
+	b.ReportAllocs()
+	b.SetBytes(int64(len(pkt)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		// IPv4 identification field is at bytes 4:6.
+		binary.BigEndian.PutUint16(pkt[4:6], uint16(i))
+		m.filterInbound(pkt, len(pkt))
+	}
+}
+
+// BenchmarkInbound_FirstFragmentDistinctIDs measures sustained first-fragment
+// pressure with distinct ids: transport decode + ACL + verdict insert until the
+// table caps, exercising the map growth and capacity guard.
+func BenchmarkInbound_FirstFragmentDistinctIDs(b *testing.B) {
+	setupBenchManager(b)
+	pkt := firstFragmentUDP(b, 0x7000, 8080, 32)
+	m := benchManager
+
+	b.ReportAllocs()
+	b.SetBytes(int64(len(pkt)))
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		binary.BigEndian.PutUint16(pkt[4:6], uint16(i))
+		m.filterInbound(pkt, len(pkt))
+	}
+}

client/firewall/uspfilter/fragment_test.go

@@ -0,0 +1,554 @@
+package uspfilter
+
+import (
+	"encoding/binary"
+	"net"
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/google/gopacket"
+	"github.com/google/gopacket/layers"
+	"github.com/stretchr/testify/require"
+
+	fw "github.com/netbirdio/netbird/client/firewall/manager"
+	nbiface "github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/iface/device"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+)
+
+const (
+	fragTestSrc   = "100.10.0.1"
+	fragTestDst   = "100.10.0.100"
+	fragTestSrcV6 = "fd00::1"
+	fragTestDstV6 = "fd00::100"
+)
+
+func newFragmentTestManager(tb testing.TB) *Manager {
+	tb.Helper()
+
+	ifaceMock := &IFaceMock{
+		SetFilterFunc: func(device.PacketFilter) error { return nil },
+		AddressFunc: func() wgaddr.Address {
+			return wgaddr.Address{
+				IP:      netip.MustParseAddr(fragTestDst),
+				Network: netip.MustParsePrefix("100.10.0.0/16"),
+				IPv6:    netip.MustParseAddr(fragTestDstV6),
+				IPv6Net: netip.MustParsePrefix("fd00::/64"),
+			}
+		},
+	}
+
+	m, err := Create(ifaceMock, false, flowLogger, nbiface.DefaultMTU)
+	require.NoError(tb, err)
+	require.NoError(tb, m.UpdateLocalIPs())
+	tb.Cleanup(func() { require.NoError(tb, m.Close(nil)) })
+	return m
+}
+
+// firstFragmentUDPTo builds the first fragment of a fragmented UDP datagram to
+// the given destination: it carries the full UDP header plus payloadLen bytes
+// of data, with the More Fragments flag set and offset zero.
+func firstFragmentUDPTo(tb testing.TB, dst string, id uint16, dstPort uint16, payloadLen int) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:  4,
+		TTL:      64,
+		Id:       id,
+		Protocol: layers.IPProtocolUDP,
+		SrcIP:    net.ParseIP(fragTestSrc),
+		DstIP:    net.ParseIP(dst),
+		Flags:    layers.IPv4MoreFragments,
+	}
+	udp := &layers.UDP{SrcPort: 40000, DstPort: layers.UDPPort(dstPort)}
+	require.NoError(tb, udp.SetNetworkLayerForChecksum(ip))
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{ComputeChecksums: true, FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, udp, gopacket.Payload(make([]byte, payloadLen))))
+	return buf.Bytes()
+}
+
+func firstFragmentUDP(tb testing.TB, id uint16, dstPort uint16, payloadLen int) []byte {
+	tb.Helper()
+	return firstFragmentUDPTo(tb, fragTestDst, id, dstPort, payloadLen)
+}
+
+// firstFragmentTCP builds the first fragment of a fragmented TCP datagram: the
+// full 20-byte TCP header plus 12 bytes of data, with the More Fragments flag
+// set and offset zero.
+func firstFragmentTCP(tb testing.TB, id uint16, dstPort uint16) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:  4,
+		TTL:      64,
+		Id:       id,
+		Protocol: layers.IPProtocolTCP,
+		SrcIP:    net.ParseIP(fragTestSrc),
+		DstIP:    net.ParseIP(fragTestDst),
+		Flags:    layers.IPv4MoreFragments,
+	}
+	tcp := &layers.TCP{SrcPort: 40000, DstPort: layers.TCPPort(dstPort), SYN: true, Window: 64240}
+	require.NoError(tb, tcp.SetNetworkLayerForChecksum(ip))
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{ComputeChecksums: true, FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, tcp, gopacket.Payload(make([]byte, 12))))
+	return buf.Bytes()
+}
+
+// trailingFragmentTo builds a non-first fragment to the given destination: an
+// IPv4 header at the given fragment offset (in 8-byte units) carrying raw
+// payload and no L4 header.
+func trailingFragmentTo(tb testing.TB, dst string, proto layers.IPProtocol, id uint16, fragOffsetOctets uint16, moreFragments bool, payloadLen int) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:    4,
+		TTL:        64,
+		Id:         id,
+		Protocol:   proto,
+		SrcIP:      net.ParseIP(fragTestSrc),
+		DstIP:      net.ParseIP(dst),
+		FragOffset: fragOffsetOctets,
+	}
+	if moreFragments {
+		ip.Flags = layers.IPv4MoreFragments
+	}
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, gopacket.Payload(make([]byte, payloadLen))))
+	return buf.Bytes()
+}
+
+func trailingFragment(tb testing.TB, id uint16, fragOffsetOctets uint16, moreFragments bool, payloadLen int) []byte {
+	tb.Helper()
+	return trailingFragmentTo(tb, fragTestDst, layers.IPProtocolUDP, id, fragOffsetOctets, moreFragments, payloadLen)
+}
+
+// outboundUDPPacket builds a complete outbound UDP packet from the local
+// address, used to establish conntrack state for reply-direction tests.
+func outboundUDPPacket(tb testing.TB, srcPort, dstPort uint16) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:  4,
+		TTL:      64,
+		Id:       1,
+		Protocol: layers.IPProtocolUDP,
+		SrcIP:    net.ParseIP(fragTestDst),
+		DstIP:    net.ParseIP(fragTestSrc),
+	}
+	udp := &layers.UDP{SrcPort: layers.UDPPort(srcPort), DstPort: layers.UDPPort(dstPort)}
+	require.NoError(tb, udp.SetNetworkLayerForChecksum(ip))
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{ComputeChecksums: true, FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, udp, gopacket.Payload(make([]byte, 16))))
+	return buf.Bytes()
+}
+
+// normalUDPPacket builds a complete, non-fragmented UDP packet for baseline
+// comparisons against the fragment paths.
+func normalUDPPacket(tb testing.TB, dstPort uint16, payloadLen int) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv4{
+		Version:  4,
+		TTL:      64,
+		Id:       1,
+		Protocol: layers.IPProtocolUDP,
+		SrcIP:    net.ParseIP(fragTestSrc),
+		DstIP:    net.ParseIP(fragTestDst),
+	}
+	udp := &layers.UDP{SrcPort: 40000, DstPort: layers.UDPPort(dstPort)}
+	require.NoError(tb, udp.SetNetworkLayerForChecksum(ip))
+
+	buf := gopacket.NewSerializeBuffer()
+	opts := gopacket.SerializeOptions{ComputeChecksums: true, FixLengths: true}
+	require.NoError(tb, gopacket.SerializeLayers(buf, opts, ip, udp, gopacket.Payload(make([]byte, payloadLen))))
+	return buf.Bytes()
+}
+
+func allowUDP(tb testing.TB, m *Manager, dstPort uint16) {
+	tb.Helper()
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrc), fw.ProtocolUDP, nil,
+		&fw.Port{Values: []uint16{dstPort}}, fw.ActionAccept, "")
+	require.NoError(tb, err)
+}
+
+// TestFragment_TrailingWithoutFirstDropped is the core bypass repro: a trailing
+// fragment with no allowed first fragment on record must be dropped. Before the
+// fix, filterInbound returned false (allow) for any fragment.
+func TestFragment_TrailingWithoutFirstDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+
+	frag := trailingFragment(t, 0x1234, 185, false, 40)
+	require.True(t, m.filterInbound(frag, len(frag)),
+		"trailing fragment without an allowed first fragment must be dropped")
+}
+
+// TestFragment_AllowedFirstPassesTrailing verifies that once a first fragment
+// passes the ACL, its trailing fragments inherit the allow verdict.
+func TestFragment_AllowedFirstPassesTrailing(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	// First fragment: UDP header (8) + 32 payload = 40 octets -> headerEnd = 5.
+	first := firstFragmentUDP(t, 0x2222, 8080, 32)
+	require.False(t, m.filterInbound(first, len(first)),
+		"allowed first fragment should pass and be recorded")
+
+	trailing := trailingFragment(t, 0x2222, 5, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of an allowed datagram should pass")
+}
+
+// TestFragment_DeniedFirstDropsTrailing verifies that a first fragment blocked
+// by the ACL leaves no verdict, so its trailing fragments are dropped.
+func TestFragment_DeniedFirstDropsTrailing(t *testing.T) {
+	m := newFragmentTestManager(t)
+	// No accept rule: local traffic defaults to deny.
+
+	first := firstFragmentUDP(t, 0x3333, 9999, 32)
+	require.True(t, m.filterInbound(first, len(first)),
+		"first fragment to a blocked port should be dropped by the ACL")
+
+	trailing := trailingFragment(t, 0x3333, 5, false, 24)
+	require.True(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of a denied datagram must be dropped")
+}
+
+// TestFragment_OverlappingHeaderDropped covers the RFC 1858 §4 / RFC 3128
+// overlapping-fragment rewrite: a trailing fragment starting inside the range
+// the ACL already inspected is dropped and poisons the datagram. TCP is used so
+// the overlap lands on real header bytes (the flags at byte 13).
+func TestFragment_OverlappingHeaderDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrc), fw.ProtocolTCP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	// First fragment: TCP header (20) + 12 data = 32 bytes -> headerEnd = 4 octets.
+	first := firstFragmentTCP(t, 0x4444, 8080)
+	require.False(t, m.filterInbound(first, len(first)))
+
+	// Overlapping fragment at offset 1 (byte 8) falls inside the inspected TCP
+	// header, so it could rewrite the flags or port on reassembly.
+	overlap := trailingFragmentTo(t, fragTestDst, layers.IPProtocolTCP, 0x4444, 1, true, 32)
+	require.True(t, m.filterInbound(overlap, len(overlap)),
+		"fragment overlapping the inspected header must be dropped")
+
+	// The datagram is now poisoned: a later, non-overlapping fragment is also
+	// dropped because the verdict was removed.
+	later := trailingFragmentTo(t, fragTestDst, layers.IPProtocolTCP, 0x4444, 4, false, 24)
+	require.True(t, m.filterInbound(later, len(later)),
+		"fragments after an overlap must be dropped (datagram poisoned)")
+}
+
+// TestFragment_OffsetZeroOverlapPoisons covers the RFC 3128 offset-zero rewrite:
+// an allowed first fragment followed by a denied offset-zero fragment for the
+// same datagram must not leave the earlier allow verdict in place.
+func TestFragment_OffsetZeroOverlapPoisons(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	allowed := firstFragmentUDP(t, 0x5A5A, 8080, 32)
+	require.False(t, m.filterInbound(allowed, len(allowed)),
+		"allowed first fragment should pass and be recorded")
+
+	// A second offset-zero fragment to a denied port supersedes the datagram's
+	// verdict; it is dropped and must not leave the allow in place.
+	denied := firstFragmentUDP(t, 0x5A5A, 9999, 32)
+	require.True(t, m.filterInbound(denied, len(denied)),
+		"denied offset-zero fragment must be dropped")
+
+	trailing := trailingFragment(t, 0x5A5A, 5, false, 24)
+	require.True(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment must be denied after the datagram was poisoned")
+}
+
+// TestFragment_TinyFirstDropped covers the tiny-fragment attack: a first
+// fragment too small to contain the full transport header can't be
+// ACL-evaluated and must be dropped.
+func TestFragment_TinyFirstDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	// IPv4 header + 4 raw bytes, MF set, offset 0: too small for the 8-byte UDP
+	// header, so it decodes to L3 only.
+	tiny := trailingFragment(t, 0x5555, 0, true, 4)
+	require.True(t, m.filterInbound(tiny, len(tiny)),
+		"tiny first fragment without a full L4 header must be dropped")
+}
+
+// TestFragment_TCPFirstFragment verifies the TCP arm of the transport decode: a
+// first fragment carrying the full 20-byte TCP header is ACL-evaluated and its
+// trailing fragments inherit the verdict.
+func TestFragment_TCPFirstFragment(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrc), fw.ProtocolTCP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	// TCP header (20) + 12 data = 32 bytes -> headerEnd = 4 octets.
+	first := firstFragmentTCP(t, 0x6666, 8080)
+	require.False(t, m.filterInbound(first, len(first)),
+		"allowed TCP first fragment should pass and be recorded")
+
+	trailing := trailingFragmentTo(t, fragTestDst, layers.IPProtocolTCP, 0x6666, 4, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of an allowed TCP datagram should pass")
+}
+
+// TestFragment_TCPTinyFirstDropped verifies the TCP minimum header length: 12
+// bytes would satisfy a UDP header but falls short of the 20-byte TCP header.
+func TestFragment_TCPTinyFirstDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrc), fw.ProtocolTCP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	tiny := trailingFragmentTo(t, fragTestDst, layers.IPProtocolTCP, 0x7777, 0, true, 12)
+	require.True(t, m.filterInbound(tiny, len(tiny)),
+		"first fragment shorter than the TCP header must be dropped")
+}
+
+// TestFragment_ConntrackAllowsFirstFragment verifies the conntrack branch: reply
+// fragments of an outbound-established UDP flow pass without any inbound rule.
+func TestFragment_ConntrackAllowsFirstFragment(t *testing.T) {
+	m := newFragmentTestManager(t)
+
+	out := outboundUDPPacket(t, 12345, 40000)
+	require.False(t, m.filterOutbound(out, len(out)))
+
+	first := firstFragmentUDP(t, 0x8888, 12345, 32)
+	require.False(t, m.filterInbound(first, len(first)),
+		"reply first fragment should pass via conntrack")
+
+	trailing := trailingFragment(t, 0x8888, 5, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of a tracked flow should pass")
+}
+
+// TestFragment_RoutingDisabledDropsFragment verifies routed first fragments are
+// dropped when routing is disabled.
+func TestFragment_RoutingDisabledDropsFragment(t *testing.T) {
+	m := newFragmentTestManager(t)
+	m.routingEnabled.Store(false)
+
+	first := firstFragmentUDPTo(t, "198.51.100.10", 0x9999, 8080, 32)
+	require.True(t, m.filterInbound(first, len(first)),
+		"routed first fragment must be dropped when routing is disabled")
+}
+
+// TestFragment_RouteACL verifies the route-ACL branch: fragments to a non-local
+// destination follow the route rules, allowed datagrams pass their trailing
+// fragments and denied ones don't.
+func TestFragment_RouteACL(t *testing.T) {
+	m := newFragmentTestManager(t)
+	m.routingEnabled.Store(true)
+	m.nativeRouter.Store(false)
+
+	_, err := m.AddRouteFiltering(
+		[]byte("rt-1"),
+		[]netip.Prefix{netip.MustParsePrefix("100.10.0.0/16")},
+		fw.Network{Prefix: netip.MustParsePrefix("198.51.100.0/24")},
+		fw.ProtocolUDP,
+		nil,
+		&fw.Port{Values: []uint16{8080}},
+		fw.ActionAccept,
+	)
+	require.NoError(t, err)
+
+	first := firstFragmentUDPTo(t, "198.51.100.10", 0xAAAA, 8080, 32)
+	require.False(t, m.filterInbound(first, len(first)),
+		"route-ACL-allowed first fragment should pass")
+	trailing := trailingFragmentTo(t, "198.51.100.10", layers.IPProtocolUDP, 0xAAAA, 5, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of an allowed routed datagram should pass")
+
+	denied := firstFragmentUDPTo(t, "198.51.100.10", 0xBBBB, 9999, 32)
+	require.True(t, m.filterInbound(denied, len(denied)),
+		"route-ACL-denied first fragment must be dropped")
+	deniedTrailing := trailingFragmentTo(t, "198.51.100.10", layers.IPProtocolUDP, 0xBBBB, 5, false, 24)
+	require.True(t, m.filterInbound(deniedTrailing, len(deniedTrailing)),
+		"trailing fragment of a denied routed datagram must be dropped")
+}
+
+// TestFragment_ExpiredVerdictDropsTrailing verifies a verdict older than the
+// tracker timeout no longer admits trailing fragments.
+func TestFragment_ExpiredVerdictDropsTrailing(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	first := firstFragmentUDP(t, 0xCCCC, 8080, 32)
+	require.False(t, m.filterInbound(first, len(first)))
+
+	m.fragments.mutex.Lock()
+	for key, entry := range m.fragments.entries {
+		entry.recordedAt = time.Now().Add(-defaultFragmentTimeout - time.Second)
+		m.fragments.entries[key] = entry
+	}
+	m.fragments.mutex.Unlock()
+
+	trailing := trailingFragment(t, 0xCCCC, 5, false, 24)
+	require.True(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment after verdict expiry must be dropped")
+}
+
+// TestFragment_CapacityFailsClosed verifies the table cap: at capacity, new
+// datagram verdicts are not recorded (their trailing fragments are dropped)
+// while already-recorded datagrams keep working.
+func TestFragment_CapacityFailsClosed(t *testing.T) {
+	m := newFragmentTestManager(t)
+	allowUDP(t, m, 8080)
+
+	m.fragments.mutex.Lock()
+	m.fragments.maxEntries = 1
+	m.fragments.mutex.Unlock()
+
+	first1 := firstFragmentUDP(t, 0x0101, 8080, 32)
+	require.False(t, m.filterInbound(first1, len(first1)))
+
+	first2 := firstFragmentUDP(t, 0x0202, 8080, 32)
+	require.False(t, m.filterInbound(first2, len(first2)),
+		"first fragment itself still passes at capacity")
+
+	trailing2 := trailingFragment(t, 0x0202, 5, false, 24)
+	require.True(t, m.filterInbound(trailing2, len(trailing2)),
+		"trailing fragment of an unrecorded datagram must be dropped at capacity")
+
+	trailing1 := trailingFragment(t, 0x0101, 5, false, 24)
+	require.False(t, m.filterInbound(trailing1, len(trailing1)),
+		"already-recorded datagram should keep passing at capacity")
+}
+
+// v6FragmentHeader builds the 8-byte IPv6 fragment extension header for the
+// given inner protocol, offset (8-byte units), More Fragments bit and id.
+func v6FragmentHeader(proto layers.IPProtocol, offsetOctets uint16, moreFragments bool, id uint32) []byte {
+	offsetFlags := offsetOctets << 3
+	if moreFragments {
+		offsetFlags |= 1
+	}
+	hdr := make([]byte, 8)
+	hdr[0] = uint8(proto)
+	binary.BigEndian.PutUint16(hdr[2:4], offsetFlags)
+	binary.BigEndian.PutUint32(hdr[4:8], id)
+	return hdr
+}
+
+func v6UDPHeader(dstPort uint16, dataLen int) []byte {
+	hdr := make([]byte, 8)
+	binary.BigEndian.PutUint16(hdr[0:2], 40000)
+	binary.BigEndian.PutUint16(hdr[2:4], dstPort)
+	binary.BigEndian.PutUint16(hdr[4:6], uint16(8+dataLen))
+	return hdr
+}
+
+// firstFragmentUDPv6 builds the first fragment of a fragmented IPv6 UDP
+// datagram: fragment header (offset 0, More Fragments set) + full UDP header +
+// data.
+func firstFragmentUDPv6(tb testing.TB, id uint32, dstPort uint16, dataLen int) []byte {
+	tb.Helper()
+	return fragmentUDPv6(tb, id, dstPort, dataLen, true)
+}
+
+// fragmentUDPv6 builds an offset-zero IPv6 UDP fragment. With moreFragments
+// false it is an atomic fragment (a complete datagram, RFC 6946).
+func fragmentUDPv6(tb testing.TB, id uint32, dstPort uint16, dataLen int, moreFragments bool) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv6{
+		Version:    6,
+		NextHeader: layers.IPProtocolIPv6Fragment,
+		HopLimit:   64,
+		SrcIP:      net.ParseIP(fragTestSrcV6),
+		DstIP:      net.ParseIP(fragTestDstV6),
+	}
+	payload := append(v6FragmentHeader(layers.IPProtocolUDP, 0, moreFragments, id), v6UDPHeader(dstPort, dataLen)...)
+	payload = append(payload, make([]byte, dataLen)...)
+
+	buf := gopacket.NewSerializeBuffer()
+	require.NoError(tb, gopacket.SerializeLayers(buf, gopacket.SerializeOptions{FixLengths: true}, ip, gopacket.Payload(payload)))
+	return buf.Bytes()
+}
+
+// trailingFragmentV6 builds a non-first IPv6 fragment: fragment header at the
+// given offset carrying raw data and no transport header.
+func trailingFragmentV6(tb testing.TB, id uint32, offsetOctets uint16, moreFragments bool, dataLen int) []byte {
+	tb.Helper()
+
+	ip := &layers.IPv6{
+		Version:    6,
+		NextHeader: layers.IPProtocolIPv6Fragment,
+		HopLimit:   64,
+		SrcIP:      net.ParseIP(fragTestSrcV6),
+		DstIP:      net.ParseIP(fragTestDstV6),
+	}
+	payload := append(v6FragmentHeader(layers.IPProtocolUDP, offsetOctets, moreFragments, id), make([]byte, dataLen)...)
+
+	buf := gopacket.NewSerializeBuffer()
+	require.NoError(tb, gopacket.SerializeLayers(buf, gopacket.SerializeOptions{FixLengths: true}, ip, gopacket.Payload(payload)))
+	return buf.Bytes()
+}
+
+// TestFragmentV6_TrailingWithoutFirstDropped verifies the IPv6 bypass is closed:
+// a trailing fragment with no allowed first fragment is dropped.
+func TestFragmentV6_TrailingWithoutFirstDropped(t *testing.T) {
+	m := newFragmentTestManager(t)
+
+	frag := trailingFragmentV6(t, 0xAABBCCDD, 100, false, 40)
+	require.True(t, m.filterInbound(frag, len(frag)),
+		"IPv6 trailing fragment without an allowed first fragment must be dropped")
+}
+
+// TestFragmentV6_AllowedFirstPassesTrailing verifies IPv6 fragments are
+// evaluated like IPv4: an allowed first fragment lets its trailing fragments
+// through.
+func TestFragmentV6_AllowedFirstPassesTrailing(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrcV6), fw.ProtocolUDP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	// First fragment: UDP header (8) + 32 data = 40 octets -> headerEnd = 5.
+	first := firstFragmentUDPv6(t, 0xAABBCCDD, 8080, 32)
+	require.False(t, m.filterInbound(first, len(first)),
+		"allowed IPv6 first fragment should pass and be recorded")
+
+	trailing := trailingFragmentV6(t, 0xAABBCCDD, 5, false, 24)
+	require.False(t, m.filterInbound(trailing, len(trailing)),
+		"trailing fragment of an allowed IPv6 datagram should pass")
+}
+
+// TestFragmentV6_AtomicNotCached verifies an IPv6 atomic fragment (fragment
+// header with offset 0 and no More Fragments, a complete datagram per RFC 6946)
+// is evaluated but not recorded, so a flood of allowed atomic fragments can't
+// exhaust the verdict table.
+func TestFragmentV6_AtomicNotCached(t *testing.T) {
+	m := newFragmentTestManager(t)
+	_, err := m.AddPeerFiltering(nil, net.ParseIP(fragTestSrcV6), fw.ProtocolUDP, nil,
+		&fw.Port{Values: []uint16{8080}}, fw.ActionAccept, "")
+	require.NoError(t, err)
+
+	atomic := fragmentUDPv6(t, 0xA70301C, 8080, 16, false)
+	require.False(t, m.filterInbound(atomic, len(atomic)),
+		"allowed IPv6 atomic fragment should pass")
+
+	m.fragments.mutex.Lock()
+	n := len(m.fragments.entries)
+	m.fragments.mutex.Unlock()
+	require.Zero(t, n, "atomic fragment must not create a verdict entry")
+
+	// A genuine fragmented datagram (More Fragments set) is still recorded.
+	first := fragmentUDPv6(t, 0xBEEF, 8080, 32, true)
+	require.False(t, m.filterInbound(first, len(first)))
+	m.fragments.mutex.Lock()
+	n = len(m.fragments.entries)
+	m.fragments.mutex.Unlock()
+	require.Equal(t, 1, n, "genuine first fragment must record a verdict")
+}

client/iface/netstack/env.go

@@ -3,14 +3,31 @@
 package netstack
 
 import (
-	"fmt"
+	"net"
 	"os"
 	"strconv"
 
 	log "github.com/sirupsen/logrus"
 )
 
-const EnvUseNetstackMode = "NB_USE_NETSTACK_MODE"
+const (
+	EnvUseNetstackMode = "NB_USE_NETSTACK_MODE"
+
+	// EnvSocks5ListenerPort overrides the port the SOCKS5 proxy listens on.
+	EnvSocks5ListenerPort = "NB_SOCKS5_LISTENER_PORT"
+
+	// EnvSocks5ListenerAddress overrides the host/IP the SOCKS5 proxy binds to.
+	// The proxy is a bridge for local host applications into the userspace
+	// WireGuard netstack, so it binds to loopback by default. Override this only
+	// when the proxy must be reachable from other hosts (e.g. a container
+	// gateway); doing so exposes an unauthenticated SOCKS5 proxy on that
+	// address.
+	EnvSocks5ListenerAddress = "NB_SOCKS5_LISTENER_ADDRESS"
+
+	// defaultSocks5Host is the loopback address the SOCKS5 proxy binds to unless
+	// overridden via EnvSocks5ListenerAddress.
+	defaultSocks5Host = "127.0.0.1"
+)
 
 // IsEnabled todo: move these function to cmd layer
 func IsEnabled() bool {
@@ -18,24 +35,40 @@ func IsEnabled() bool {
 }
 
 func ListenAddr() string {
-	sPort := os.Getenv("NB_SOCKS5_LISTENER_PORT")
+	return net.JoinHostPort(listenHost(), strconv.Itoa(listenPort()))
+}
+
+// listenHost returns the host/IP the SOCKS5 proxy binds to. It defaults to
+// loopback and only honors EnvSocks5ListenerAddress when it holds a valid IP.
+func listenHost() string {
+	addr := os.Getenv(EnvSocks5ListenerAddress)
+	if addr == "" {
+		return defaultSocks5Host
+	}
+	if net.ParseIP(addr) == nil {
+		log.Warnf("invalid socks5 listener address %q, falling back to default: %s", addr, defaultSocks5Host)
+		return defaultSocks5Host
+	}
+	return addr
+}
+
+// listenPort returns the port the SOCKS5 proxy binds to, defaulting to
+// DefaultSocks5Port when EnvSocks5ListenerPort is unset or invalid.
+func listenPort() int {
+	sPort := os.Getenv(EnvSocks5ListenerPort)
 	if sPort == "" {
-		return listenAddr(DefaultSocks5Port)
+		return DefaultSocks5Port
 	}
 
 	port, err := strconv.Atoi(sPort)
 	if err != nil {
 		log.Warnf("invalid socks5 listener port, unable to convert it to int, falling back to default: %d", DefaultSocks5Port)
-		return listenAddr(DefaultSocks5Port)
+		return DefaultSocks5Port
 	}
 	if port < 1 || port > 65535 {
 		log.Warnf("invalid socks5 listener port, it should be in the range 1-65535, falling back to default: %d", DefaultSocks5Port)
-		return listenAddr(DefaultSocks5Port)
+		return DefaultSocks5Port
 	}
 
-	return listenAddr(port)
-}
-
-func listenAddr(port int) string {
-	return fmt.Sprintf("0.0.0.0:%d", port)
+	return port
 }

client/iface/netstack/env_test.go

@@ -0,0 +1,63 @@
+//go:build !js
+
+package netstack
+
+import (
+	"net"
+	"strconv"
+	"testing"
+)
+
+func TestListenAddr_DefaultsToLoopback(t *testing.T) {
+	// No env overrides: must bind loopback, never all interfaces.
+	got := ListenAddr()
+	want := net.JoinHostPort("127.0.0.1", strconv.Itoa(DefaultSocks5Port))
+	if got != want {
+		t.Fatalf("ListenAddr() = %q, want %q", got, want)
+	}
+}
+
+func TestListenAddr_AddressOverride(t *testing.T) {
+	tests := []struct {
+		name string
+		env  string
+		want string
+	}{
+		{name: "valid override honored", env: "0.0.0.0", want: "0.0.0.0"},
+		{name: "valid specific ip honored", env: "10.0.0.5", want: "10.0.0.5"},
+		{name: "ipv6 loopback bracketed", env: "::1", want: "::1"},
+		{name: "invalid falls back to loopback", env: "not-an-ip", want: "127.0.0.1"},
+		{name: "empty falls back to loopback", env: "", want: "127.0.0.1"},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv(EnvSocks5ListenerAddress, tc.env)
+			want := net.JoinHostPort(tc.want, strconv.Itoa(DefaultSocks5Port))
+			if got := ListenAddr(); got != want {
+				t.Fatalf("ListenAddr() = %q, want %q", got, want)
+			}
+		})
+	}
+}
+
+func TestListenAddr_PortOverride(t *testing.T) {
+	tests := []struct {
+		name string
+		env  string
+		want int
+	}{
+		{name: "valid port honored", env: "1081", want: 1081},
+		{name: "non-numeric falls back", env: "abc", want: DefaultSocks5Port},
+		{name: "out of range falls back", env: "70000", want: DefaultSocks5Port},
+		{name: "zero falls back", env: "0", want: DefaultSocks5Port},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv(EnvSocks5ListenerPort, tc.env)
+			want := net.JoinHostPort("127.0.0.1", strconv.Itoa(tc.want))
+			if got := ListenAddr(); got != want {
+				t.Fatalf("ListenAddr() = %q, want %q", got, want)
+			}
+		})
+	}
+}

client/internal/auth/device_flow.go

@@ -299,7 +299,7 @@ func (d *DeviceAuthorizationFlow) WaitToken(ctx context.Context, info AuthFlowIn
 				UseIDToken:   d.providerConfig.UseIDToken,
 			}
 
-			err = isValidAccessToken(tokenInfo.GetTokenToUse(), d.providerConfig.Audience)
+			err = validateTokenAudience(tokenInfo.GetTokenToUse(), d.providerConfig.Audience)
 			if err != nil {
 				return TokenInfo{}, fmt.Errorf("validate access token failed with error: %v", err)
 			}

client/internal/auth/pkce_flow.go

@@ -306,7 +306,7 @@ func (p *PKCEAuthorizationFlow) parseOAuthToken(token *oauth2.Token) (TokenInfo,
 		audience = p.providerConfig.ClientID
 	}
 
-	if err := isValidAccessToken(tokenInfo.GetTokenToUse(), audience); err != nil {
+	if err := validateTokenAudience(tokenInfo.GetTokenToUse(), audience); err != nil {
 		return TokenInfo{}, fmt.Errorf("authentication failed: invalid access token - %w", err)
 	}
 
@@ -320,6 +320,11 @@ func (p *PKCEAuthorizationFlow) parseOAuthToken(token *oauth2.Token) (TokenInfo,
 	return tokenInfo, nil
 }
 
+// parseEmailFromIDToken extracts the email (or name) claim from an ID token
+// without verifying its signature. The value is best-effort and used only as a
+// UX convenience (login hint prefill and display); it never drives an
+// authorization decision. The authoritative identity is established server-side
+// from the signature-verified token.
 func parseEmailFromIDToken(token string) (string, error) {
 	parts := strings.Split(token, ".")
 	if len(parts) < 2 {

client/internal/auth/util.go

@@ -20,14 +20,26 @@ func randomBytesInHex(count int) (string, error) {
 	return hex.EncodeToString(buf), nil
 }
 
-// isValidAccessToken is a simple validation of the access token
-func isValidAccessToken(token string, audience string) error {
+// validateTokenAudience checks that the token is a well-formed JWT whose
+// audience claim matches the expected audience.
+//
+// It does NOT verify the token's cryptographic signature and therefore must not
+// be treated as an authenticity check. The token is obtained by the client
+// directly from the IdP token endpoint over TLS, and its signature is verified
+// server-side by the management server against the IdP's JWKS
+// (see shared/auth/jwt/validator.go). This function is only a client-side
+// sanity check that the returned token targets the expected audience.
+func validateTokenAudience(token string, audience string) error {
 	if token == "" {
 		return fmt.Errorf("token received is empty")
 	}
 
-	encodedClaims := strings.Split(token, ".")[1]
-	claimsString, err := base64.RawURLEncoding.DecodeString(encodedClaims)
+	parts := strings.Split(token, ".")
+	if len(parts) != 3 {
+		return fmt.Errorf("token is not a well-formed JWT")
+	}
+
+	claimsString, err := base64.RawURLEncoding.DecodeString(parts[1])
 	if err != nil {
 		return err
 	}

client/internal/auth/util_test.go

@@ -0,0 +1,108 @@
+package auth
+
+import (
+	"encoding/base64"
+	"encoding/json"
+	"testing"
+)
+
+// makeJWT builds an unsigned JWT-shaped string (header.payload.signature) with
+// the given claims payload. The signature part is arbitrary because
+// validateTokenAudience intentionally does not verify it.
+func makeJWT(t *testing.T, claims map[string]interface{}) string {
+	t.Helper()
+	header := base64.RawURLEncoding.EncodeToString([]byte(`{"alg":"RS256","typ":"JWT"}`))
+	payloadBytes, err := json.Marshal(claims)
+	if err != nil {
+		t.Fatalf("marshal claims: %v", err)
+	}
+	payload := base64.RawURLEncoding.EncodeToString(payloadBytes)
+	return header + "." + payload + ".unverified-signature"
+}
+
+func TestValidateTokenAudience(t *testing.T) {
+	tests := []struct {
+		name     string
+		token    string
+		audience string
+		wantErr  bool
+	}{
+		{
+			name:     "empty token",
+			token:    "",
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "not a JWT - no dots",
+			token:    "notajwt",
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "not a JWT - two parts only",
+			token:    "header.payload",
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "matching string audience",
+			token:    makeJWT(t, map[string]interface{}{"aud": "netbird"}),
+			audience: "netbird",
+			wantErr:  false,
+		},
+		{
+			name:     "mismatching string audience",
+			token:    makeJWT(t, map[string]interface{}{"aud": "other"}),
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "matching audience in array",
+			token:    makeJWT(t, map[string]interface{}{"aud": []interface{}{"other", "netbird"}}),
+			audience: "netbird",
+			wantErr:  false,
+		},
+		{
+			name:     "mismatching audience array",
+			token:    makeJWT(t, map[string]interface{}{"aud": []interface{}{"a", "b"}}),
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "missing audience claim",
+			token:    makeJWT(t, map[string]interface{}{"sub": "user"}),
+			audience: "netbird",
+			wantErr:  true,
+		},
+		{
+			name:     "invalid base64 payload",
+			token:    "header.!!!not-base64!!!.sig",
+			audience: "netbird",
+			wantErr:  true,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			err := validateTokenAudience(tc.token, tc.audience)
+			if tc.wantErr && err == nil {
+				t.Fatalf("expected error, got nil")
+			}
+			if !tc.wantErr && err != nil {
+				t.Fatalf("expected no error, got %v", err)
+			}
+		})
+	}
+}
+
+// TestValidateTokenAudienceNoPanic guards the regression where a non-empty
+// token without the JWT dot structure caused an index-out-of-range panic.
+func TestValidateTokenAudienceNoPanic(t *testing.T) {
+	inputs := []string{"a", ".", "a.", "aaaa", "no-dots-here"}
+	for _, in := range inputs {
+		if err := validateTokenAudience(in, "netbird"); err == nil {
+			t.Fatalf("expected error for malformed token %q, got nil", in)
+		}
+	}
+}

client/internal/debug/debug.go

@@ -480,7 +480,6 @@ func (g *BundleGenerator) addStatus() error {
 
 		fullStatus := g.statusRecorder.GetFullStatus()
 		protoFullStatus := nbstatus.ToProtoFullStatus(fullStatus)
-		protoFullStatus.Events = g.statusRecorder.GetEventHistory()
 		overview := nbstatus.ConvertToStatusOutputOverview(protoFullStatus, nbstatus.ConvertOptions{
 			Anonymize:     g.anonymize,
 			ProfileName:   profName,

client/internal/dns/service_listener.go

@@ -292,18 +292,16 @@ func (s *serviceViaListener) generateFreePort() (uint16, error) {
 		return customPort, nil
 	}
 
-	udpAddr := net.UDPAddrFromAddrPort(netip.MustParseAddrPort("0.0.0.0:0"))
-	probeListener, err := net.ListenUDP("udp", udpAddr)
+	probeListener, err := net.ListenUDP("udp4", &net.UDPAddr{})
 	if err != nil {
 		log.Debugf("failed to bind random port for DNS: %s", err)
 		return 0, err
 	}
 
-	addrPort := netip.MustParseAddrPort(probeListener.LocalAddr().String()) // might panic if address is incorrect
-	err = probeListener.Close()
-	if err != nil {
+	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 addrPort.Port(), nil
+	return port, nil
 }

client/internal/engine.go

@@ -2605,13 +2605,14 @@ func (e *Engine) updateForwardRules(rules []*mgmProto.ForwardingRule) ([]firewal
 
 func (e *Engine) toExcludedLazyPeers(rules []firewallManager.ForwardRule, 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 {
-		ip := r.TranslatedAddress
 		for _, p := range peers {
-			for _, allowedIP := range p.GetAllowedIps() {
-				if allowedIP != ip.String() {
-					continue
-				}
+			if e.peerRoutesAddr(p, r.TranslatedAddress) {
 				log.Infof("exclude forwarder peer from lazy connection: %s", p.GetWgPubKey())
 				excludedPeers[p.GetWgPubKey()] = true
 			}
@@ -2621,6 +2622,27 @@ func (e *Engine) toExcludedLazyPeers(rules []firewallManager.ForwardRule, peers
 	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_lazy_exclude_test.go

@@ -0,0 +1,87 @@
+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/peer/conn.go

@@ -203,7 +203,6 @@ func NewConn(config ConnConfig, services ServiceDependencies) (*Conn, error) {
 		statusICE:          worker.NewAtomicStatus(),
 		dumpState:          dumpState,
 		endpointUpdater:    NewEndpointUpdater(connLog, config.WgConfig, isController(config)),
-		wgWatcher:          NewWGWatcher(connLog, config.WgConfig.WgInterface, config.Key, dumpState),
 		metricsRecorder:    services.MetricsRecorder,
 	}
 
@@ -671,11 +670,12 @@ func (conn *Conn) onGuardEvent() {
 	}
 }
 
-func (conn *Conn) onWGDisconnected() {
+func (conn *Conn) onWGDisconnected(watcherCtx context.Context) {
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
 
-	if conn.ctx.Err() != nil {
+	// watcherCtx guards against a stale watcher tearing down a connection that already superseded it.
+	if conn.ctx.Err() != nil || watcherCtx.Err() != nil {
 		return
 	}
 
@@ -833,25 +833,39 @@ func (conn *Conn) isConnectedOnAllWay() (status guard.ConnStatus) {
 	})
 }
 
+// enableWgWatcherIfNeeded starts a fresh watcher instance per connection attempt, so its
+// lifecycle stays bound to conn.mu and enable/disable can't race an old goroutine's shutdown.
+// Caller must hold conn.mu.
 func (conn *Conn) enableWgWatcherIfNeeded(enabledTime time.Time) {
-	if !conn.wgWatcher.PrepareInitialHandshake() {
+	if conn.wgWatcher != nil {
 		return
 	}
 
+	watcher := NewWGWatcher(conn.Log, conn.config.WgConfig.WgInterface, conn.config.Key, conn.dumpState)
+	watcher.PrepareInitialHandshake()
+
 	wgWatcherCtx, wgWatcherCancel := context.WithCancel(conn.ctx)
+	conn.wgWatcher = watcher
 	conn.wgWatcherCancel = wgWatcherCancel
+
 	conn.wgWatcherWg.Add(1)
 	go func() {
 		defer conn.wgWatcherWg.Done()
-		conn.wgWatcher.EnableWgWatcher(wgWatcherCtx, enabledTime, conn.onWGDisconnected, conn.onWGHandshakeSuccess, conn.onWGCheckSuccess)
+		onDisconnected := func() { conn.onWGDisconnected(wgWatcherCtx) }
+		watcher.EnableWgWatcher(wgWatcherCtx, enabledTime, onDisconnected, conn.onWGHandshakeSuccess, conn.onWGCheckSuccess)
 	}()
 }
 
+// disableWgWatcherIfNeeded cancels and drops the watcher once no transport is active. It never
+// waits for the goroutine: the timeout path reentrantly calls back here under conn.mu, so
+// blocking would deadlock. Caller must hold conn.mu.
 func (conn *Conn) disableWgWatcherIfNeeded() {
-	if conn.currentConnPriority == conntype.None && conn.wgWatcherCancel != nil {
-		conn.wgWatcherCancel()
-		conn.wgWatcherCancel = nil
+	if conn.currentConnPriority != conntype.None || conn.wgWatcher == nil {
+		return
 	}
+	conn.wgWatcherCancel()
+	conn.wgWatcher = nil
+	conn.wgWatcherCancel = nil
 }
 
 func (conn *Conn) newProxy(remoteConn net.Conn) (wgproxy.Proxy, error) {
@@ -874,7 +888,9 @@ func (conn *Conn) resetEndpoint() {
 		return
 	}
 	conn.Log.Infof("reset wg endpoint")
-	conn.wgWatcher.Reset()
+	if conn.wgWatcher != nil {
+		conn.wgWatcher.Reset()
+	}
 	if err := conn.endpointUpdater.RemoveEndpointAddress(); err != nil {
 		conn.Log.Warnf("failed to remove endpoint address before update: %v", err)
 	}

client/internal/peer/conn_test.go

@@ -339,20 +339,20 @@ func TestConn_onWGDisconnected_EscalatesToRosenpassReset(t *testing.T) {
 	conn := newWGTimeoutTestConn(true, &disconnected)
 
 	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	assert.Empty(t, disconnected, "escalation must not fire below the threshold")
 
-	conn.onWGDisconnected()
+	conn.onWGDisconnected(conn.ctx)
 	assert.Equal(t, []string{conn.config.WgConfig.RemoteKey}, disconnected,
 		"reaching the threshold must report the peer disconnected once")
 
 	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	assert.Len(t, disconnected, 1, "escalation must restart counting after firing")
 
-	conn.onWGDisconnected()
+	conn.onWGDisconnected(conn.ctx)
 	assert.Len(t, disconnected, 2, "continued timeouts must escalate again")
 }
 
@@ -364,12 +364,12 @@ func TestConn_onWGDisconnected_CheckSuccessResetsEscalation(t *testing.T) {
 	conn := newWGTimeoutTestConn(true, &disconnected)
 
 	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	conn.onWGCheckSuccess()
 
 	for i := 0; i < wgTimeoutEscalationThreshold-1; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	assert.Empty(t, disconnected, "handshake success must reset the timeout count")
 }
@@ -382,7 +382,7 @@ func TestConn_onWGDisconnected_NoEscalationWithoutRosenpass(t *testing.T) {
 	conn := newWGTimeoutTestConn(false, &disconnected)
 
 	for i := 0; i < wgTimeoutEscalationThreshold*3; i++ {
-		conn.onWGDisconnected()
+		conn.onWGDisconnected(conn.ctx)
 	}
 	assert.Empty(t, disconnected, "escalation must be limited to rosenpass connections")
 }

client/internal/peer/wg_watcher.go

@@ -3,7 +3,6 @@ package peer
 import (
 	"context"
 	"fmt"
-	"sync"
 	"time"
 
 	log "github.com/sirupsen/logrus"
@@ -24,14 +23,14 @@ type WGInterfaceStater interface {
 	GetStats() (map[string]configurer.WGStats, error)
 }
 
+// WGWatcher is single-shot: one instance per connection attempt, run once, then discarded.
+// Lifecycle is owned by Conn under conn.mu, so it keeps no "enabled" state to go stale.
 type WGWatcher struct {
 	log           *log.Entry
 	wgIfaceStater WGInterfaceStater
 	peerKey       string
 	stateDump     *stateDump
 
-	enabled   bool
-	muEnabled sync.Mutex
 	// initialHandshake is not thread-safe; never call PrepareInitialHandshake and EnableWgWatcher concurrently.
 	initialHandshake time.Time
 
@@ -48,25 +47,14 @@ func NewWGWatcher(log *log.Entry, wgIfaceStater WGInterfaceStater, peerKey strin
 	}
 }
 
-// PrepareInitialHandshake reserves the watcher and reads the peer's current WireGuard
-// handshake time. It must be called before the peer is (re)configured on the WireGuard
-// interface, so the captured baseline reflects the state prior to this connection attempt
-// instead of racing with that configuration. Returns ok=false if the watcher is already
-// running, in which case EnableWgWatcher must not be called.
-func (w *WGWatcher) PrepareInitialHandshake() (ok bool) {
-	w.muEnabled.Lock()
-	if w.enabled {
-		w.muEnabled.Unlock()
-		return false
-	}
-
+// PrepareInitialHandshake reads the peer's current WireGuard handshake time. It must be
+// called before the peer is (re)configured on the WireGuard interface, so the captured
+// baseline reflects the state prior to this connection attempt instead of racing with
+// that configuration.
+func (w *WGWatcher) PrepareInitialHandshake() {
 	w.log.Debugf("enable WireGuard watcher")
-	w.enabled = true
-	w.muEnabled.Unlock()
-
 	handshake, _ := w.wgState()
 	w.initialHandshake = handshake
-	return true
 }
 
 // EnableWgWatcher runs the WireGuard watcher loop using the handshake baseline captured by
@@ -76,10 +64,6 @@ func (w *WGWatcher) PrepareInitialHandshake() (ok bool) {
 // handshake, including the first.
 func (w *WGWatcher) EnableWgWatcher(ctx context.Context, enabledTime time.Time, onDisconnectedFn func(), onHandshakeSuccessFn func(when time.Time), onCheckSuccessFn func()) {
 	w.periodicHandshakeCheck(ctx, onDisconnectedFn, onHandshakeSuccessFn, onCheckSuccessFn, enabledTime, w.initialHandshake)
-
-	w.muEnabled.Lock()
-	w.enabled = false
-	w.muEnabled.Unlock()
 }
 
 // Reset signals the watcher that the WireGuard peer has been reset and a new
@@ -105,6 +89,7 @@ func (w *WGWatcher) periodicHandshakeCheck(ctx context.Context, onDisconnectedFn
 		case <-timer.C:
 			handshake, ok := w.handshakeCheck(lastHandshake)
 			if !ok {
+				// early ctx cancel check return
 				if ctx.Err() != nil {
 					return
 				}
@@ -153,9 +138,9 @@ func (w *WGWatcher) handshakeCheck(lastHandshake time.Time) (*time.Time, bool) {
 
 	w.log.Tracef("previous handshake, handshake: %v, %v", lastHandshake, handshake)
 
-	// the current know handshake did not change
+	// the current known handshake did not change
 	if handshake.Equal(lastHandshake) {
-		w.log.Warnf("WireGuard handshake timed out: %v", handshake)
+		w.log.Warnf("WireGuard handshake not updated: %v", handshake)
 		return nil, false
 	}
 

client/internal/peer/wg_watcher_test.go

@@ -7,7 +7,6 @@ import (
 	"time"
 
 	log "github.com/sirupsen/logrus"
-	"github.com/stretchr/testify/require"
 
 	"github.com/netbirdio/netbird/client/iface/configurer"
 )
@@ -62,7 +61,7 @@ func TestWGWatcher_CheckSuccessCallback(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
-	require.True(t, watcher.PrepareInitialHandshake())
+	watcher.PrepareInitialHandshake()
 
 	firstHandshake := make(chan struct{}, 1)
 	checkSuccess := make(chan struct{}, 1)
@@ -101,8 +100,7 @@ func TestWGWatcher_EnableWgWatcher(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	defer cancel()
 
-	ok := watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should not be enabled yet")
+	watcher.PrepareInitialHandshake()
 
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {
@@ -132,8 +130,7 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	watcher := NewWGWatcher(mlog, mocWgIface, "", newStateDump("peer", mlog, &Status{}))
 
 	ctx, cancel := context.WithCancel(context.Background())
-	ok := watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should not be enabled yet")
+	watcher.PrepareInitialHandshake()
 
 	wg := &sync.WaitGroup{}
 	wg.Add(1)
@@ -149,8 +146,7 @@ func TestWGWatcher_ReEnable(t *testing.T) {
 	ctx, cancel = context.WithCancel(context.Background())
 	defer cancel()
 
-	ok = watcher.PrepareInitialHandshake()
-	require.True(t, ok, "watcher should be re-enabled after the previous run stopped")
+	watcher.PrepareInitialHandshake()
 
 	onDisconnected := make(chan struct{}, 1)
 	go watcher.EnableWgWatcher(ctx, time.Now(), func() {

client/internal/statemanager/manager.go

@@ -1,6 +1,7 @@
 package statemanager
 
 import (
+	"bytes"
 	"context"
 	"encoding/json"
 	"errors"
@@ -305,6 +306,11 @@ func (m *Manager) loadStateFile(deleteCorrupt bool) (map[string]json.RawMessage,
 
 	var rawStates map[string]json.RawMessage
 	if err := json.Unmarshal(data, &rawStates); err != nil {
+		if len(bytes.TrimSpace(data)) == 0 {
+			log.Warnf("state file %s is empty (%d bytes)", m.filePath, len(data))
+		} else {
+			log.Warnf("state file %s has malformed content (%d bytes)", m.filePath, len(data))
+		}
 		m.handleCorruptedState(deleteCorrupt)
 		return nil, fmt.Errorf("unmarshal states: %w", err)
 	}

client/ios/NetBirdSDK/client.go

@@ -233,6 +233,9 @@ func (c *Client) DebugBundle(anonymize bool) (string, error) {
 		deps.SyncResponse = resp
 
 		if e := cc.Engine(); e != nil {
+			deps.RefreshStatus = func() {
+				e.RunHealthProbes(context.Background(), true)
+			}
 			if cm := e.GetClientMetrics(); cm != nil {
 				deps.ClientMetrics = cm
 			}

client/ios/NetBirdSDK/login.go

@@ -44,10 +44,25 @@ type Auth struct {
 // NewAuth instantiate Auth struct and validate the management URL
 func NewAuth(cfgPath string, mgmURL string) (*Auth, error) {
 	inputCfg := profilemanager.ConfigInput{
+		ConfigPath:    cfgPath,
 		ManagementURL: mgmURL,
 	}
 
-	cfg, err := profilemanager.CreateInMemoryConfig(inputCfg)
+	// Load the existing config when a config file is already present so an
+	// interactive re-login reuses the peer's persisted WireGuard private key
+	// (and thus its identity) instead of generating a fresh one. Generating a
+	// new key registers a brand-new peer on the management server on every
+	// re-auth (named after the fallback hostname). Only fall back to a fresh
+	// in-memory config for the first-time login when no config file exists yet.
+	// DirectUpdateOrCreateConfig uses non-atomic writes so it also works inside
+	// the tvOS App Group sandbox where atomic temp-file+rename is blocked.
+	var cfg *profilemanager.Config
+	var err error
+	if cfgPath != "" {
+		cfg, err = profilemanager.DirectUpdateOrCreateConfig(inputCfg)
+	} else {
+		cfg, err = profilemanager.CreateInMemoryConfig(inputCfg)
+	}
 	if err != nil {
 		return nil, err
 	}

client/mdm/canonical_loaders.go

@@ -22,6 +22,7 @@ var allKeys = []string{
 	KeyDisableMetricsCollection,
 	KeyAllowServerSSH,
 	KeyDisableAutoConnect,
+	KeyDisableAutostart,
 	KeyPreSharedKey,
 	KeyRosenpassEnabled,
 	KeyRosenpassPermissive,

client/mdm/policy.go

@@ -20,10 +20,10 @@ import (
 // names (lowerCamelCase) so the daemon can map a Policy key directly to a
 // configuration field.
 const (
-	KeyManagementURL            = "managementURL"
-	KeyDisableUpdateSettings    = "disableUpdateSettings"
-	KeyDisableProfiles          = "disableProfiles"
-	KeyDisableNetworks          = "disableNetworks"
+	KeyManagementURL         = "managementURL"
+	KeyDisableUpdateSettings = "disableUpdateSettings"
+	KeyDisableProfiles       = "disableProfiles"
+	KeyDisableNetworks       = "disableNetworks"
 	// KeyDisableAdvancedView gates the advanced-view section in the
 	// upcoming UI revision. UI-only: NOT stored on Config, not
 	// applied by applyMDMPolicy, not rejectable via SetConfig. The
@@ -37,10 +37,16 @@ const (
 	KeyDisableMetricsCollection = "disableMetricsCollection"
 	KeyAllowServerSSH           = "allowServerSSH"
 	KeyDisableAutoConnect       = "disableAutoConnect"
-	KeyPreSharedKey             = "preSharedKey"
-	KeyRosenpassEnabled         = "rosenpassEnabled"
-	KeyRosenpassPermissive      = "rosenpassPermissive"
-	KeyWireguardPort            = "wireguardPort"
+	// KeyDisableAutostart suppresses the GUI's fresh-install
+	// launch-on-login default and marks the Settings toggle as
+	// MDM-managed. UI-only: NOT stored on Config and not applied by
+	// applyMDMPolicy; the GUI reads it directly and it appears in
+	// GetConfigResponse.mDMManagedFields when set.
+	KeyDisableAutostart    = "disableAutostart"
+	KeyPreSharedKey        = "preSharedKey"
+	KeyRosenpassEnabled    = "rosenpassEnabled"
+	KeyRosenpassPermissive = "rosenpassPermissive"
+	KeyWireguardPort       = "wireguardPort"
 
 	// Split tunnel is modeled as a single conceptual policy with two
 	// registry/plist values. KeySplitTunnelMode is the discriminator

client/server/server.go

@@ -181,7 +181,7 @@ func (s *Server) Start() error {
 		log.Warnf("failed to redirect stderr: %v", err)
 	}
 
-	if err := restoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
+	if err := RestoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
 		log.Warnf(errRestoreResidualState, err)
 	}
 
@@ -551,7 +551,7 @@ func (s *Server) Login(callerCtx context.Context, msg *proto.LoginRequest) (*pro
 	s.actCancel = cancel
 	s.mutex.Unlock()
 
-	if err := restoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
+	if err := RestoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
 		log.Warnf(errRestoreResidualState, err)
 	}
 
@@ -858,7 +858,7 @@ func (s *Server) Up(callerCtx context.Context, msg *proto.UpRequest) (*proto.UpR
 
 		return s.waitForUp(callerCtx)
 	}
-	if err := restoreResidualState(callerCtx, s.profileManager.GetStatePath()); err != nil {
+	if err := RestoreResidualState(callerCtx, s.profileManager.GetStatePath()); err != nil {
 		log.Warnf(errRestoreResidualState, err)
 	}
 

client/server/state.go

@@ -46,7 +46,7 @@ func (s *Server) CleanState(ctx context.Context, req *proto.CleanStateRequest) (
 
 	if req.All {
 		// Reuse existing cleanup logic for all states
-		if err := restoreResidualState(ctx, statePath); err != nil {
+		if err := RestoreResidualState(ctx, statePath); err != nil {
 			return nil, status.Errorf(codes.Internal, "failed to clean all states: %v", err)
 		}
 
@@ -113,9 +113,9 @@ func (s *Server) DeleteState(ctx context.Context, req *proto.DeleteStateRequest)
 	}, nil
 }
 
-// restoreResidualState checks if the client was not shut down in a clean way and restores residual if required.
+// RestoreResidualState checks if the client was not shut down in a clean way and restores residual if required.
 // Otherwise, we might not be able to connect to the management server to retrieve new config.
-func restoreResidualState(ctx context.Context, statePath string) error {
+func RestoreResidualState(ctx context.Context, statePath string) error {
 	if statePath == "" {
 		return nil
 	}

client/ssh/config/manager.go

@@ -14,6 +14,7 @@ import (
 	log "github.com/sirupsen/logrus"
 
 	nbssh "github.com/netbirdio/netbird/client/ssh"
+	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
 const (
@@ -218,11 +219,20 @@ func (m *Manager) buildHostPatterns(peer PeerSSHInfo) []string {
 	if peer.IPv6.IsValid() {
 		hostPatterns = append(hostPatterns, peer.IPv6.String())
 	}
-	if peer.FQDN != "" {
+	// Peer FQDNs and hostnames originate from remote peers, so they must be
+	// validated as plain DNS names before being embedded in the ssh_config
+	// "Match host" pattern list. This prevents injection of arbitrary
+	// ssh_config directives via embedded quotes, whitespace, newlines, the
+	// comma pattern separator, or the "*"/"?" pattern metacharacters.
+	if domain.IsValidDomainNoWildcard(peer.FQDN) {
 		hostPatterns = append(hostPatterns, peer.FQDN)
+	} else if peer.FQDN != "" {
+		log.Warnf("skipping peer FQDN with invalid characters in SSH config: %q", peer.FQDN)
 	}
-	if peer.Hostname != "" && peer.Hostname != peer.FQDN {
+	if peer.Hostname != peer.FQDN && domain.IsValidDomainNoWildcard(peer.Hostname) {
 		hostPatterns = append(hostPatterns, peer.Hostname)
+	} else if peer.Hostname != "" && peer.Hostname != peer.FQDN {
+		log.Warnf("skipping peer hostname with invalid characters in SSH config: %q", peer.Hostname)
 	}
 	return hostPatterns
 }

client/ssh/config/manager_test.go

@@ -148,6 +148,45 @@ func TestManager_MatchHostFormat(t *testing.T) {
 		"should use Match host with comma-separated patterns")
 }
 
+func TestManager_HostPatternInjection(t *testing.T) {
+	tempDir, err := os.MkdirTemp("", "netbird-ssh-config-test")
+	require.NoError(t, err)
+	defer func() { assert.NoError(t, os.RemoveAll(tempDir)) }()
+
+	manager := &Manager{
+		sshConfigDir:  filepath.Join(tempDir, "ssh_config.d"),
+		sshConfigFile: "99-netbird.conf",
+	}
+
+	// A malicious peer FQDN/hostname attempts to break out of the Match host
+	// directive and inject arbitrary ssh_config (a ProxyCommand executing a
+	// command). It must be rejected, not written to the config.
+	peers := []PeerSSHInfo{
+		{
+			Hostname: "evil\"\n    ProxyCommand touch /tmp/pwned\nHost x",
+			IP:       netip.MustParseAddr("100.125.1.1"),
+			FQDN:     "evil\"\n    ProxyCommand touch /tmp/pwned\nHost x.nb.internal",
+		},
+		{Hostname: "peer2", IP: netip.MustParseAddr("100.125.1.2"), FQDN: "peer2.nb.internal"},
+	}
+
+	err = manager.SetupSSHClientConfig(peers)
+	require.NoError(t, err)
+
+	configPath := filepath.Join(manager.sshConfigDir, manager.sshConfigFile)
+	content, err := os.ReadFile(configPath)
+	require.NoError(t, err)
+	configStr := string(content)
+
+	assert.NotContains(t, configStr, "ProxyCommand touch /tmp/pwned",
+		"injected directive must not appear in generated config")
+	assert.NotContains(t, configStr, "evil",
+		"malicious pattern must be dropped entirely")
+	// The valid peer must still be present, on a single Match host line.
+	assert.Contains(t, configStr, "Match host \"100.125.1.1,100.125.1.2,peer2.nb.internal,peer2\"",
+		"valid peers must survive, injected patterns dropped")
+}
+
 func TestManager_ForcedSSHConfig(t *testing.T) {
 	// Set force environment variable
 	t.Setenv(EnvForceSSHConfig, "true")

client/ssh/server/getent_unix.go

@@ -69,7 +69,8 @@ func parseGetentPasswd(output string) (*user.User, string, error) {
 
 // validateGetentInput checks that the input is safe to pass to getent or id.
 // Allows POSIX usernames, numeric UIDs, and common NSS extensions
-// (@ for Kerberos, $ for Samba, + for NIS compat).
+// (@ 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 validateGetentInput(input string) bool {
 	maxLen := 32
 	if runtime.GOOS == "linux" {
@@ -80,6 +81,10 @@ func validateGetentInput(input string) bool {
 		return false
 	}
 
+	if input[0] == '-' {
+		return false
+	}
+
 	for _, r := range input {
 		if isAllowedGetentChar(r) {
 			continue

client/ssh/server/getent_unix_test.go

@@ -157,6 +157,9 @@ func TestValidateGetentInput(t *testing.T) {
 		{"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},

client/status/status.go

@@ -746,6 +746,8 @@ func ToProtoFullStatus(fullStatus peer.FullStatus) *proto.FullStatus {
 		pbFullStatus.DnsServers = append(pbFullStatus.DnsServers, pbDnsState)
 	}
 
+	pbFullStatus.Events = fullStatus.Events
+
 	return &pbFullStatus
 }
 

client/ui/autostart_default.go

@@ -0,0 +1,107 @@
+//go:build !android && !ios && !freebsd && !js
+
+package main
+
+import (
+	"context"
+	"os"
+	"path/filepath"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/profilemanager"
+	"github.com/netbirdio/netbird/client/mdm"
+	"github.com/netbirdio/netbird/client/ui/preferences"
+	"github.com/netbirdio/netbird/client/ui/services"
+)
+
+// autostartDefaultState carries the guard inputs of the one-time autostart
+// default decision so the decision itself stays a pure, testable function.
+type autostartDefaultState struct {
+	supported    bool
+	mdmDisabled  bool
+	priorInstall bool
+}
+
+// shouldEnableAutostartDefault applies the first-run guards in order and
+// returns whether autostart may be enabled, plus the reason when it may not.
+func shouldEnableAutostartDefault(s autostartDefaultState) (bool, string) {
+	switch {
+	case !s.supported:
+		return false, "autostart not supported on this platform"
+	case s.mdmDisabled:
+		return false, "autostart disabled by MDM policy"
+	case s.priorInstall:
+		return false, "existing NetBird installation"
+	}
+	return true, ""
+}
+
+// autostartDisabledByMDM reports whether the MDM policy manages the
+// disableAutostart key in a way that must suppress the default. An
+// unparseable managed value is treated as disabled to stay on the safe side.
+func autostartDisabledByMDM(policy *mdm.Policy) bool {
+	if !policy.HasKey(mdm.KeyDisableAutostart) {
+		return false
+	}
+	disabled, ok := policy.GetBool(mdm.KeyDisableAutostart)
+	return !ok || disabled
+}
+
+// netbirdFootprintExists reports whether the machine already carries NetBird
+// daemon config or state, meaning this is not a genuinely fresh install. It is
+// the update-safety gate for the autostart default: upgrading users always
+// have a footprint, so an update can never trigger a login-item write.
+func netbirdFootprintExists() bool {
+	candidates := []string{
+		profilemanager.DefaultConfigPath,
+		filepath.Join(profilemanager.DefaultConfigPathDir, "config.json"),
+		filepath.Join(profilemanager.DefaultConfigPathDir, "state.json"),
+	}
+	for _, path := range candidates {
+		if path != "" && fileExists(path) {
+			return true
+		}
+	}
+	return false
+}
+
+// applyAutostartDefault runs the one-time launch-on-login default for genuinely
+// fresh installs. The autostartInitialized marker is persisted before any
+// enable attempt so a crash mid-flow degrades to "never enabled" instead of
+// retrying login-item writes on every launch. A user's later disable in
+// Settings is never overridden: the marker guarantees at-most-once, ever.
+func applyAutostartDefault(ctx context.Context, autostart *services.Autostart, prefs *preferences.Store, prefsFileExisted bool) {
+	priorFootprint := netbirdFootprintExists() || prefsFileExisted
+
+	if prefs.Get().AutostartInitialized {
+		return
+	}
+	if err := prefs.SetAutostartInitialized(true); err != nil {
+		log.Warnf("persist autostart marker, skipping autostart default: %v", err)
+		return
+	}
+
+	state := autostartDefaultState{
+		supported:    autostart.Supported(ctx),
+		mdmDisabled:  autostartDisabledByMDM(mdm.LoadPolicy()),
+		priorInstall: priorFootprint,
+	}
+	enable, reason := shouldEnableAutostartDefault(state)
+	if !enable {
+		log.Debugf("skipping autostart default: %s", reason)
+		return
+	}
+
+	if err := autostart.SetEnabled(ctx, true); err != nil {
+		log.Warnf("enable autostart on fresh install: %v", err)
+		return
+	}
+	log.Info("autostart enabled by default on fresh install")
+}
+
+// fileExists reports whether path exists.
+func fileExists(path string) bool {
+	_, err := os.Stat(path)
+	return err == nil
+}

client/ui/autostart_default_test.go

@@ -0,0 +1,125 @@
+//go:build !android && !ios && !freebsd && !js
+
+package main
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+
+	"github.com/netbirdio/netbird/client/mdm"
+)
+
+func TestShouldEnableAutostartDefault(t *testing.T) {
+	allPass := autostartDefaultState{
+		supported:    true,
+		mdmDisabled:  false,
+		priorInstall: false,
+	}
+
+	tests := []struct {
+		name       string
+		mutate     func(*autostartDefaultState)
+		wantEnable bool
+		wantReason string
+	}{
+		{
+			name:       "fresh install with all guards passing enables",
+			mutate:     func(*autostartDefaultState) {},
+			wantEnable: true,
+		},
+		{
+			name:       "unsupported platform skips",
+			mutate:     func(s *autostartDefaultState) { s.supported = false },
+			wantReason: "autostart not supported on this platform",
+		},
+		{
+			name:       "MDM disable skips",
+			mutate:     func(s *autostartDefaultState) { s.mdmDisabled = true },
+			wantReason: "autostart disabled by MDM policy",
+		},
+		{
+			name:       "existing installation (upgrade) skips",
+			mutate:     func(s *autostartDefaultState) { s.priorInstall = true },
+			wantReason: "existing NetBird installation",
+		},
+		{
+			name: "unsupported wins over every other guard",
+			mutate: func(s *autostartDefaultState) {
+				s.supported = false
+				s.mdmDisabled = true
+				s.priorInstall = true
+			},
+			wantReason: "autostart not supported on this platform",
+		},
+		{
+			name: "MDM disable wins over prior install",
+			mutate: func(s *autostartDefaultState) {
+				s.mdmDisabled = true
+				s.priorInstall = true
+			},
+			wantReason: "autostart disabled by MDM policy",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			state := allPass
+			tc.mutate(&state)
+			enable, reason := shouldEnableAutostartDefault(state)
+			assert.Equal(t, tc.wantEnable, enable, "enable decision should match for state %+v", state)
+			assert.Equal(t, tc.wantReason, reason, "skip reason should identify the failing guard")
+		})
+	}
+}
+
+func TestAutostartDisabledByMDM(t *testing.T) {
+	tests := []struct {
+		name   string
+		values map[string]any
+		want   bool
+	}{
+		{
+			name:   "empty policy does not disable",
+			values: nil,
+			want:   false,
+		},
+		{
+			name:   "unrelated managed keys do not disable",
+			values: map[string]any{mdm.KeyDisableAutoConnect: true},
+			want:   false,
+		},
+		{
+			name:   "disableAutostart true disables",
+			values: map[string]any{mdm.KeyDisableAutostart: true},
+			want:   true,
+		},
+		{
+			name:   "disableAutostart registry DWORD 1 disables",
+			values: map[string]any{mdm.KeyDisableAutostart: int64(1)},
+			want:   true,
+		},
+		{
+			name:   "disableAutostart string true disables",
+			values: map[string]any{mdm.KeyDisableAutostart: "true"},
+			want:   true,
+		},
+		{
+			name:   "disableAutostart explicit false allows",
+			values: map[string]any{mdm.KeyDisableAutostart: false},
+			want:   false,
+		},
+		{
+			name:   "unparseable managed value is treated as disabled",
+			values: map[string]any{mdm.KeyDisableAutostart: "not-a-bool"},
+			want:   true,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got := autostartDisabledByMDM(mdm.NewPolicy(tc.values))
+			assert.Equal(t, tc.want, got, "MDM disable decision should match for values %v", tc.values)
+		})
+	}
+}

client/ui/frontend/src/lib/connection.ts

@@ -51,7 +51,14 @@ async function runSsoLogin(
     if (uri) await openBrowserLoginUri(uri);
 
     const cancelPromise = buildSsoCancelPromise(state, signal);
-    const waitPromise = Connection.WaitSSOLogin({ userCode: result.userCode, hostname: "" });
+    // Combine wait + up in Go so the connection comes up the moment SSO
+    // completes. During SSO the tray window is hidden and the webview is
+    // suspended, so a frontend-driven Up (a promise continuation) would not
+    // fire until the user woke the window (e.g. hovering the tray icon).
+    const waitPromise = Connection.WaitSSOLoginAndUp(
+        { userCode: result.userCode, hostname: "" },
+        { profileName: "", username: "" },
+    );
 
     try {
         await Promise.race([waitPromise, cancelPromise]);
@@ -89,13 +96,13 @@ export async function startConnection(onSettled?: () => void, signal?: AbortSign
         if (signal?.aborted) state.cancelled = true;
 
         if (!state.cancelled && result.needsSsoLogin) {
+            // runSsoLogin brings the connection up in Go once SSO completes.
             await runSsoLogin(result, state, signal);
-        }
-
-        if (!state.cancelled && signal?.aborted) state.cancelled = true;
-
-        if (!state.cancelled) {
-            await Connection.Up({ profileName: "", username: "" });
+        } else {
+            if (!state.cancelled && signal?.aborted) state.cancelled = true;
+            if (!state.cancelled) {
+                await Connection.Up({ profileName: "", username: "" });
+            }
         }
     } catch (e) {
         WindowManager.CloseBrowserLogin().catch(console.error);

client/ui/i18n/locales/_index.json

@@ -8,6 +8,7 @@
         {"code": "fr", "displayName": "Français", "englishName": "French"},
         {"code": "it", "displayName": "Italiano", "englishName": "Italian"},
         {"code": "pt", "displayName": "Português", "englishName": "Portuguese"},
-        {"code": "zh-CN", "displayName": "简体中文", "englishName": "Simplified Chinese"}
+        {"code": "zh-CN", "displayName": "简体中文", "englishName": "Simplified Chinese"},
+        {"code": "ja", "displayName": "日本語", "englishName": "Japanese"}
     ]
 }

client/ui/main.go

@@ -197,6 +197,9 @@ func main() {
 		// daemon may keep the main window from showing, so the OS toast is the
 		// only reliable signal the user gets.
 		go notifyIfDaemonOutdated(compat, notifier, localizer)
+		// One-time launch-on-login default for fresh installs; gated by the
+		// NetBird footprint check, MDM policy, and the persisted marker.
+		go applyAutostartDefault(context.Background(), services.NewAutostart(app.Autostart), prefStore, prefStore.ExistedAtLoad())
 	})
 
 	if err := app.Run(); err != nil {

client/ui/preferences/store.go

@@ -54,6 +54,10 @@ type UIPreferences struct {
 	Language            i18n.LanguageCode `json:"language"`
 	ViewMode            ViewMode          `json:"viewMode"`
 	OnboardingCompleted bool              `json:"onboardingCompleted"`
+	// AutostartInitialized records that the one-time autostart default
+	// decision has run for this OS user. It only ever transitions to true
+	// and is never reset, so the default-on flow runs at most once, ever.
+	AutostartInitialized bool `json:"autostartInitialized"`
 }
 
 // LanguageValidator rejects SetLanguage inputs with no shipped bundle.
@@ -72,8 +76,9 @@ type Emitter interface {
 type Store struct {
 	path string
 
-	mu      sync.RWMutex
-	current UIPreferences
+	mu            sync.RWMutex
+	current       UIPreferences
+	existedAtLoad bool
 
 	subsMu sync.Mutex
 	subs   []chan UIPreferences
@@ -157,6 +162,27 @@ func (s *Store) SetOnboardingCompleted(done bool) error {
 	return nil
 }
 
+// SetAutostartInitialized persists the one-time autostart decision marker.
+// No-op if unchanged.
+func (s *Store) SetAutostartInitialized(done bool) error {
+	s.mu.Lock()
+	if s.current.AutostartInitialized == done {
+		s.mu.Unlock()
+		return nil
+	}
+	next := s.current
+	next.AutostartInitialized = done
+	if err := s.persistLocked(next); err != nil {
+		s.mu.Unlock()
+		return fmt.Errorf("persist preferences: %w", err)
+	}
+	s.current = next
+	s.mu.Unlock()
+
+	s.broadcast(next)
+	return nil
+}
+
 // SetLanguage validates, persists, and broadcasts. No-op if unchanged.
 func (s *Store) SetLanguage(lang i18n.LanguageCode) error {
 	if lang == "" {
@@ -206,13 +232,29 @@ func (s *Store) Subscribe() (<-chan UIPreferences, func()) {
 	return ch, unsubscribe
 }
 
+// ExistedAtLoad reports whether the backing preferences file was present on
+// disk when the store loaded. It distinguishes a user who ran a prior GUI
+// version from a brand-new OS user with no preferences yet.
+func (s *Store) ExistedAtLoad() bool {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	return s.existedAtLoad
+}
+
 // load reads the file into current. A missing file is not an error (the
 // in-memory default stands); malformed contents return an error.
 func (s *Store) load() error {
-	if _, err := os.Stat(s.path); errors.Is(err, os.ErrNotExist) {
-		return nil
+	if _, err := os.Stat(s.path); err != nil {
+		if errors.Is(err, os.ErrNotExist) {
+			return nil
+		}
+		return fmt.Errorf("stat preferences: %w", err)
 	}
 
+	s.mu.Lock()
+	s.existedAtLoad = true
+	s.mu.Unlock()
+
 	var loaded UIPreferences
 	if _, err := util.ReadJson(s.path, &loaded); err != nil {
 		return err

client/ui/preferences/store_test.go

@@ -215,6 +215,46 @@ func TestStore_FileShapeIsJSON(t *testing.T) {
 	assert.Equal(t, i18n.LanguageCode("hu"), parsed.Language)
 }
 
+func TestStore_SetAutostartInitializedPersistsAcrossReload(t *testing.T) {
+	withTempConfigDir(t)
+	emitter := &recordingEmitter{}
+	s, err := NewStore(nil, emitter)
+	require.NoError(t, err)
+
+	assert.False(t, s.Get().AutostartInitialized, "marker must default to false when no file is on disk")
+
+	require.NoError(t, s.SetAutostartInitialized(true))
+	assert.True(t, s.Get().AutostartInitialized, "Get should reflect the persisted marker")
+	require.Len(t, emitter.calledWith(EventPreferencesChanged), 1, "first marker write should broadcast")
+
+	// Re-setting the same value must be a no-op: no disk write, no broadcast.
+	require.NoError(t, s.SetAutostartInitialized(true))
+	assert.Len(t, emitter.calledWith(EventPreferencesChanged), 1, "idempotent marker write should not broadcast again")
+
+	// A fresh Store (new GUI launch) must see the marker so the autostart
+	// default decision never runs twice.
+	reloaded, err := NewStore(nil, nil)
+	require.NoError(t, err)
+	assert.True(t, reloaded.Get().AutostartInitialized, "marker must survive a reload from disk")
+}
+
+func TestStore_ExistedAtLoad(t *testing.T) {
+	withTempConfigDir(t)
+
+	// Brand-new OS user: no preferences file on disk yet.
+	fresh, err := NewStore(nil, nil)
+	require.NoError(t, err)
+	assert.False(t, fresh.ExistedAtLoad(), "ExistedAtLoad must be false when no file is on disk")
+
+	// Persisting a value writes the file to disk.
+	require.NoError(t, fresh.SetLanguage("en"))
+
+	// A subsequent GUI launch reopens the now-present file.
+	reopened, err := NewStore(nil, nil)
+	require.NoError(t, err)
+	assert.True(t, reopened.ExistedAtLoad(), "ExistedAtLoad must be true after the store has persisted and is reopened")
+}
+
 func TestStore_ErrUnsupportedSentinel(t *testing.T) {
 	// Verifies callers can match on the sentinel error rather than parsing
 	// strings — protects against accidental %v -> %w changes that would

client/ui/services/connection.go

@@ -35,7 +35,7 @@ type LoginResult struct {
 	VerificationURIComplete string `json:"verificationUriComplete"`
 }
 
-// WaitSSOParams are the inputs to WaitSSOLogin.
+// WaitSSOParams are the inputs to waitSSOLogin.
 type WaitSSOParams struct {
 	UserCode string `json:"userCode"`
 	Hostname string `json:"hostname"`
@@ -125,23 +125,6 @@ func (s *Connection) Login(ctx context.Context, p LoginParams) (LoginResult, err
 	}, nil
 }
 
-func (s *Connection) WaitSSOLogin(ctx context.Context, p WaitSSOParams) (string, error) {
-	cli, err := s.conn.Client()
-	if err != nil {
-		return "", err
-	}
-	log.Infof("waiting for SSO login to complete")
-	resp, err := cli.WaitSSOLogin(ctx, &proto.WaitSSOLoginRequest{
-		UserCode: p.UserCode,
-		Hostname: p.Hostname,
-	})
-	if err != nil {
-		return "", s.classifyDaemonError(err)
-	}
-	log.Infof("SSO login completed, daemon reported success")
-	return resp.GetEmail(), nil
-}
-
 func (s *Connection) Up(ctx context.Context, p UpParams) error {
 	cli, err := s.conn.Client()
 	if err != nil {
@@ -162,6 +145,27 @@ func (s *Connection) Up(ctx context.Context, p UpParams) error {
 	return nil
 }
 
+// WaitSSOLoginAndUp blocks until the SSO login completes and then brings the
+// connection up, both from the Go side. Keeping the post-login Up here rather
+// than as a frontend continuation is deliberate: during SSO the tray window is
+// hidden and the webview is suspended (macOS App Nap / hidden-window timer
+// throttling), so a frontend-driven Up would not run until the user woke the
+// window (e.g. by hovering the tray icon). Doing it in Go connects the moment
+// the daemon reports SSO success. Returns the authenticated user's email.
+func (s *Connection) WaitSSOLoginAndUp(ctx context.Context, wait WaitSSOParams, up UpParams) (string, error) {
+	email, err := s.waitSSOLogin(ctx, wait)
+	if err != nil {
+		return "", err
+	}
+	if err := ctx.Err(); err != nil {
+		return "", err
+	}
+	if err := s.Up(ctx, up); err != nil {
+		return "", err
+	}
+	return email, nil
+}
+
 func (s *Connection) Down(ctx context.Context) error {
 	cli, err := s.conn.Client()
 	if err != nil {
@@ -221,6 +225,26 @@ func (s *Connection) Logout(ctx context.Context, p LogoutParams) error {
 	return nil
 }
 
+// waitSSOLogin blocks until the daemon reports the SSO login result and returns
+// the authenticated user's email. It is unexported because the frontend drives
+// SSO through the exported WaitSSOLoginAndUp.
+func (s *Connection) waitSSOLogin(ctx context.Context, p WaitSSOParams) (string, error) {
+	cli, err := s.conn.Client()
+	if err != nil {
+		return "", err
+	}
+	log.Infof("waiting for SSO login to complete")
+	resp, err := cli.WaitSSOLogin(ctx, &proto.WaitSSOLoginRequest{
+		UserCode: p.UserCode,
+		Hostname: p.Hostname,
+	})
+	if err != nil {
+		return "", s.classifyDaemonError(err)
+	}
+	log.Infof("SSO login completed, daemon reported success")
+	return resp.GetEmail(), nil
+}
+
 // classifyDaemonError maps a gRPC error to a localised ClientError.
 func (s *Connection) classifyDaemonError(err error) *ClientError {
 	return s.classifier.classify(err)

client/ui/services/settings.go

@@ -20,11 +20,12 @@ type MDMFields struct {
 	DisableServerRoutes      bool   `json:"disableServerRoutes"`
 	AllowServerSSH           *bool  `json:"allowServerSSH"`
 	DisableAutoConnect       bool   `json:"disableAutoConnect"`
+	DisableAutostart         bool   `json:"disableAutostart"`
 	BlockInbound             bool   `json:"blockInbound"`
 	DisableMetricsCollection bool   `json:"disableMetricsCollection"`
 	SplitTunnelMode          bool   `json:"splitTunnelMode"`
 	SplitTunnelApps          bool   `json:"splitTunnelApps"`
-	DisableAdvancedView bool `json:"disableAdvancedView"`
+	DisableAdvancedView      bool   `json:"disableAdvancedView"`
 }
 
 type Features struct {

combined/cmd/config.go

@@ -145,6 +145,7 @@ type AuthConfig struct {
 	CLIRedirectURIs                 []string          `yaml:"cliRedirectURIs"`
 	Owner                           *AuthOwnerConfig  `yaml:"owner,omitempty"`
 	DashboardPostLogoutRedirectURIs []string          `yaml:"dashboardPostLogoutRedirectURIs"`
+	GrantTypes                      []string          `yaml:"grantTypes"`
 }
 
 // AuthStorageConfig contains auth storage settings
@@ -604,6 +605,7 @@ func (c *CombinedConfig) buildEmbeddedIdPConfig(mgmt ManagementConfig) (*idp.Emb
 		DashboardRedirectURIs:           mgmt.Auth.DashboardRedirectURIs,
 		CLIRedirectURIs:                 mgmt.Auth.CLIRedirectURIs,
 		DashboardPostLogoutRedirectURIs: mgmt.Auth.DashboardPostLogoutRedirectURIs,
+		GrantTypes:                      mgmt.Auth.GrantTypes,
 	}
 
 	if mgmt.Auth.Owner != nil && mgmt.Auth.Owner.Email != "" {

combined/cmd/root.go

@@ -226,7 +226,7 @@ func (s *serverInstances) createRelayServer(cfg *CombinedConfig, tlsSupport bool
 	}
 
 	hashedSecret := sha256.Sum256([]byte(cfg.Relay.AuthSecret))
-	authenticator := auth.NewTimedHMACValidator(hashedSecret[:], 24*time.Hour)
+	authenticator := auth.NewTimedHMACValidator(hashedSecret[:])
 
 	relayCfg := relayServer.Config{
 		Meter:          s.metricsServer.Meter,

e2e/agentnetwork/chat_test.go

@@ -39,6 +39,21 @@ func availableProviders() []providerCase {
 	if k := os.Getenv("ANTHROPIC_TOKEN"); k != "" {
 		ps = append(ps, providerCase{name: "anthropic", catalogID: "anthropic_api", upstream: "https://api.anthropic.com", apiKey: k, model: "claude-haiku-4-5", kind: harness.WireMessages})
 	}
+	if k := os.Getenv("KIMI_TOKEN"); k != "" {
+		// Kimi (Moonshot AI) serves two body shapes from the same key: OpenAI
+		// Chat Completions on the bare host (/v1/...) and the Anthropic
+		// Messages API under the /anthropic path prefix (the endpoint
+		// Moonshot's Claude Code guide uses). The platform serves this
+		// account exactly ONE model — kimi-k3 (kimi-k2-thinking and even
+		// kimi-latest return resource_not_found_error on both surfaces) — so
+		// two concurrent provider records would claim the same model and
+		// route ambiguously. Run the Anthropic shape, the flagship Claude
+		// Code path; the OpenAI wire shape is covered live by the other
+		// chat-shaped matrix providers, and Kimi-over-chat passed with
+		// kimi-k3 before the single-model constraint surfaced (run #73 on
+		// the kimi feature branch).
+		ps = append(ps, providerCase{name: "kimi", catalogID: "kimi_api", upstream: "https://api.moonshot.ai/anthropic", apiKey: k, model: "kimi-k3", kind: harness.WireMessages})
+	}
 	if k, u := os.Getenv("VERCEL_TOKEN"), os.Getenv("VERCEL_URL"); k != "" && u != "" {
 		ps = append(ps, providerCase{name: "vercel", catalogID: "vercel_ai_gateway", upstream: u, apiKey: k, model: "openai/gpt-4o-mini", kind: harness.WireChat})
 	}
@@ -84,18 +99,38 @@ func availableProviders() []providerCase {
 		}
 	}
 
-	// Bedrock: path-routed, bearer auth. Model is a cross-region inference
-	// profile id (distinct string from the first-party Anthropic case).
+	// Bedrock: path-routed, bearer auth. Model is the FULL cross-region
+	// inference-profile id exactly as AWS issues it — region-family prefix
+	// plus the date/version suffix. A bare or wrong-region id makes Bedrock
+	// reject the request with "The provided model identifier is invalid"
+	// before any inference runs. The proxy normalizes this id to the catalog
+	// key (anthropic.claude-haiku-4-5) for routing/pricing/allowlists.
+	// Defaults pair eu-central-1 with the eu.* profile; AWS_REGION overrides
+	// the region and the prefix follows its family.
 	if k := os.Getenv("AWS_BEARER_TOKEN_BEDROCK"); k != "" {
 		region := os.Getenv("AWS_REGION")
 		if region == "" {
-			region = "us-east-1"
+			region = "eu-central-1"
 		}
-		ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: "us.anthropic.claude-haiku-4-5", kind: harness.WireMessages})
+		ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: bedrockProfilePrefix(region) + ".anthropic.claude-haiku-4-5-20251001-v1:0", kind: harness.WireBedrock})
 	}
 	return ps
 }
 
+// bedrockProfilePrefix maps an AWS region to its cross-region inference
+// profile prefix: us-east-1 -> us, eu-central-1 -> eu, ap-southeast-1 -> apac,
+// us-gov-west-1 -> us-gov.
+func bedrockProfilePrefix(region string) string {
+	switch {
+	case strings.HasPrefix(region, "us-gov-"):
+		return "us-gov"
+	case strings.HasPrefix(region, "ap-"):
+		return "apac"
+	default:
+		return strings.SplitN(region, "-", 2)[0]
+	}
+}
+
 // providerRequest builds a create request for a matrix provider: enabled, with
 // a uniquely-priced model for body-routed providers and none for the
 // path-routed Vertex (whose model lives in the request path).
@@ -224,9 +259,12 @@ func TestProvidersMatrix(t *testing.T) {
 				var c int
 				var b string
 				var cerr error
-				if pc.kind == harness.WireVertex {
+				switch pc.kind {
+				case harness.WireVertex:
 					c, b, cerr = cl.Vertex(ctx, settings.Endpoint, proxyIP, pc.project, pc.region, pc.model, "Reply with exactly: pong", sessionID)
-				} else {
+				case harness.WireBedrock:
+					c, b, cerr = cl.Bedrock(ctx, settings.Endpoint, proxyIP, pc.model, "Reply with exactly: pong", sessionID)
+				default:
 					c, b, cerr = cl.Chat(ctx, settings.Endpoint, proxyIP, pc.kind, pc.model, "Reply with exactly: pong", sessionID)
 				}
 				if cerr == nil {

e2e/agentnetwork/guardrail_test.go

@@ -0,0 +1,187 @@
+//go:build e2e
+
+package agentnetwork
+
+import (
+	"context"
+	"regexp"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/e2e/harness"
+	"github.com/netbirdio/netbird/shared/management/http/api"
+)
+
+// bedrockVersionSuffix mirrors the proxy's normalizer: the trailing
+// "-vN[:N]" or "-YYYYMMDD-vN[:N]" version/throughput suffix of a Bedrock
+// model id.
+var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
+
+// catalogModel returns the normalized catalog id the proxy stamps for a
+// path-routed provider's configured model — the form the guardrail allowlist is
+// compared against (region prefix / @version stripped).
+func catalogModel(pc providerCase) string {
+	switch pc.kind {
+	case harness.WireBedrock:
+		// Mirror the proxy's llm.NormalizeBedrockModel (not importable from
+		// here — proxy/internal): strip the region-family prefix and the
+		// date/version suffix so the allowlist carries the catalog key the
+		// guardrail compares against.
+		m := pc.model
+		for _, p := range []string{"us.", "eu.", "apac.", "global."} {
+			if strings.HasPrefix(m, p) {
+				m = strings.TrimPrefix(m, p)
+				break
+			}
+		}
+		return bedrockVersionSuffix.ReplaceAllString(m, "")
+	case harness.WireVertex:
+		return strings.SplitN(pc.model, "@", 2)[0]
+	default:
+		return pc.model
+	}
+}
+
+// disallowedModel returns a valid-shaped model id for the provider that is NOT
+// the configured/allowed one, so the guardrail must reject it before the
+// request ever reaches the upstream.
+func disallowedModel(pc providerCase) string {
+	switch pc.kind {
+	case harness.WireBedrock:
+		// Same profile prefix as the allowed model so only the model name
+		// differs; the guardrail must deny it before it reaches AWS.
+		return strings.SplitN(pc.model, ".", 2)[0] + ".anthropic.claude-opus-4-8"
+	case harness.WireVertex:
+		return "claude-opus-4-8@20250101"
+	default:
+		return "unlisted-model"
+	}
+}
+
+// sendModel drives one request for the given model through the provider's native
+// wire shape and returns the HTTP status.
+func sendModel(ctx context.Context, t *testing.T, cl *harness.Client, endpoint, proxyIP string, pc providerCase, model string) int {
+	t.Helper()
+	var code int
+	var err error
+	switch pc.kind {
+	case harness.WireBedrock:
+		code, _, err = cl.Bedrock(ctx, endpoint, proxyIP, model, "Reply with exactly: pong", "")
+	case harness.WireVertex:
+		code, _, err = cl.Vertex(ctx, endpoint, proxyIP, pc.project, pc.region, model, "Reply with exactly: pong", "")
+	default:
+		code, _, err = cl.Chat(ctx, endpoint, proxyIP, pc.kind, model, "Reply with exactly: pong", "")
+	}
+	require.NoError(t, err, "request must reach the proxy for %s", pc.name)
+	return code
+}
+
+// TestModelAllowlistEnforced provisions a Model Allowlist guardrail limiting each
+// path-routed provider (Bedrock, Vertex) to its configured model, then drives
+// requests over the tunnel: the allowed model returns 200 while a model outside
+// the allowlist is denied 403 by the guardrail before it reaches the upstream.
+// This is the coverage missing for #6751 — the model for these providers travels
+// in the URL path, and the allowlist must be enforced there.
+func TestModelAllowlistEnforced(t *testing.T) {
+	var providers []providerCase
+	for _, pc := range availableProviders() {
+		if pc.kind == harness.WireBedrock || pc.kind == harness.WireVertex {
+			providers = append(providers, pc)
+		}
+	}
+	if len(providers) == 0 {
+		t.Skip("no path-routed provider keys set (AWS_BEARER_TOKEN_BEDROCK / GOOGLE_VERTEX_*); source ~/.llm-keys")
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Minute)
+	defer cancel()
+
+	grp, err := srv.API().Groups.Create(ctx, api.PostApiGroupsJSONRequestBody{Name: "e2e-allowlist"})
+	require.NoError(t, err, "create group")
+	t.Cleanup(func() { _ = srv.API().Groups.Delete(context.Background(), grp.Id) })
+
+	ephemeral := false
+	sk, err := srv.API().SetupKeys.Create(ctx, api.PostApiSetupKeysJSONRequestBody{
+		Name:       "e2e-allowlist-client",
+		Type:       "reusable",
+		ExpiresIn:  86400,
+		UsageLimit: 0,
+		AutoGroups: []string{grp.Id},
+		Ephemeral:  &ephemeral,
+	})
+	require.NoError(t, err, "mint setup key")
+
+	// Providers with their configured (allowed) models; the first bootstraps the cluster.
+	ids := make([]string, 0, len(providers))
+	allowed := make([]string, 0, len(providers))
+	for i, pc := range providers {
+		req := providerRequest(pc)
+		if i == 0 {
+			req.BootstrapCluster = ptr(harness.AgentNetworkCluster)
+		}
+		prov, perr := srv.CreateProvider(ctx, req)
+		require.NoError(t, perr, "create provider %s", pc.name)
+		id := prov.Id
+		ids = append(ids, id)
+		allowed = append(allowed, catalogModel(pc))
+		t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), id) })
+	}
+
+	// Guardrail allowlisting exactly the configured models.
+	var gr api.AgentNetworkGuardrailRequest
+	gr.Name = "e2e-allowlist"
+	gr.Checks.ModelAllowlist.Enabled = true
+	gr.Checks.ModelAllowlist.Models = allowed
+	guard, err := srv.CreateGuardrail(ctx, gr)
+	require.NoError(t, err, "create guardrail")
+	t.Cleanup(func() { _ = srv.DeleteGuardrail(context.Background(), guard.Id) })
+
+	enabled := true
+	pol, err := srv.CreatePolicy(ctx, api.AgentNetworkPolicyRequest{
+		Name:                   "e2e-allowlist",
+		Enabled:                &enabled,
+		SourceGroups:           []string{grp.Id},
+		DestinationProviderIds: ids,
+		GuardrailIds:           &[]string{guard.Id},
+	})
+	require.NoError(t, err, "create policy")
+	t.Cleanup(func() { _ = srv.DeletePolicy(context.Background(), pol.Id) })
+
+	settings, err := srv.GetSettings(ctx)
+	require.NoError(t, err, "read settings for endpoint")
+	require.NotEmpty(t, settings.Endpoint, "agent-network endpoint must be assigned")
+
+	proxyToken, err := srv.CreateProxyTokenCLI(ctx, "e2e-proxy-allowlist")
+	require.NoError(t, err, "mint proxy token via CLI")
+	px, err := harness.StartProxy(ctx, srv, proxyToken)
+	require.NoError(t, err, "start proxy")
+	t.Cleanup(func() { _ = px.Terminate(context.Background()) })
+
+	cl, err := harness.StartClient(ctx, srv, sk.Key)
+	require.NoError(t, err, "start client")
+	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
+
+	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	if err := cl.WaitProxyPeer(ctx, 180*time.Second); err != nil {
+		t.Fatalf("client did not see the proxy peer: %v\n=== proxy logs ===\n%s", err, px.Logs(context.Background()))
+	}
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve agent-network endpoint to proxy IP")
+
+	for _, pc := range providers {
+		pc := pc
+		t.Run(pc.name, func(t *testing.T) {
+			// The admin's allowlisted model is served end to end.
+			assert.Equal(t, 200, sendModel(ctx, t, cl, settings.Endpoint, proxyIP, pc, pc.model),
+				"allowlisted model must be permitted for %s", pc.name)
+			// A model outside the allowlist is rejected by the guardrail (before
+			// the upstream), regardless of whether it is a real catalog model.
+			assert.Equal(t, 403, sendModel(ctx, t, cl, settings.Endpoint, proxyIP, pc, disallowedModel(pc)),
+				"model outside the allowlist must be denied for %s", pc.name)
+		})
+	}
+}

e2e/harness/agentnetwork.go

@@ -107,6 +107,17 @@ func (c *Combined) DeletePolicy(ctx context.Context, id string) error {
 	return anDelete(ctx, c, "/api/agent-network/policies/"+id)
 }
 
+// CreateGuardrail creates an agent-network guardrail (e.g. a model allowlist)
+// that can then be attached to a policy via its GuardrailIds.
+func (c *Combined) CreateGuardrail(ctx context.Context, req api.AgentNetworkGuardrailRequest) (api.AgentNetworkGuardrail, error) {
+	return anRequest[api.AgentNetworkGuardrail](ctx, c, http.MethodPost, "/api/agent-network/guardrails", req)
+}
+
+// DeleteGuardrail removes a guardrail by id.
+func (c *Combined) DeleteGuardrail(ctx context.Context, id string) error {
+	return anDelete(ctx, c, "/api/agent-network/guardrails/"+id)
+}
+
 // GetSettings returns the account's agent-network settings row. It exists only
 // after the first provider create bootstraps it.
 func (c *Combined) GetSettings(ctx context.Context) (api.AgentNetworkSettings, error) {

e2e/harness/client.go

@@ -56,6 +56,11 @@ func StartClient(ctx context.Context, c *Combined, setupKey string) (*Client, er
 			// Match the proxy: the combined relay is WebSocket-only, so the
 			// client must use WS transport to keep a stable relay link to it.
 			"NB_RELAY_TRANSPORT": "ws",
+			// Lazy connections defer the peer dial until traffic flows, which
+			// leaves `netbird status` at "0/N Connected" and starves
+			// WaitProxyPeer. Force eager connections so the proxy peer shows
+			// Connected as soon as the tunnel is up.
+			"NB_LAZY_CONN": "false",
 		},
 		HostConfigModifier: func(hc *container.HostConfig) {
 			hc.CapAdd = append(hc.CapAdd, "NET_ADMIN", "SYS_ADMIN", "SYS_RESOURCE")
@@ -194,6 +199,11 @@ const (
 	// WireVertex is the Anthropic-on-Vertex rawPredict shape: the client posts
 	// the full Vertex model path and the proxy mints the SA OAuth token.
 	WireVertex = "vertex"
+	// WireBedrock is the native AWS Bedrock InvokeModel shape: the model id
+	// travels in the URL path (/model/{id}/invoke), not the body, so the proxy
+	// routes by path. This is what a Bedrock SDK client sends and the shape the
+	// model-allowlist guardrail must enforce.
+	WireBedrock = "bedrock"
 )
 
 // Chat issues a chat-completion POST to the agent-network endpoint over the
@@ -226,6 +236,17 @@ func (cl *Client) Vertex(ctx context.Context, endpoint, proxyIP, project, region
 	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
 }
 
+// Bedrock issues a native AWS Bedrock InvokeModel POST over the tunnel. The
+// model id is carried in the request path (/model/{id}/invoke), so the proxy
+// routes by path; the body uses the bedrock anthropic_version rather than a
+// model field. A non-empty sessionID is sent as the universal x-session-id
+// header the proxy records.
+func (cl *Client) Bedrock(ctx context.Context, endpoint, proxyIP, model, prompt, sessionID string) (int, string, error) {
+	path := "/model/" + model + "/invoke"
+	body := fmt.Sprintf(`{"anthropic_version":"bedrock-2023-05-31","max_tokens":64,"messages":[{"role":"user","content":%q}]}`, prompt)
+	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
+}
+
 // withSessionID appends the x-session-id header when sessionID is non-empty.
 func withSessionID(headers []string, sessionID string) []string {
 	if sessionID == "" {

e2e/harness/proxy.go

@@ -77,6 +77,10 @@ func StartProxy(ctx context.Context, c *Combined, proxyToken string) (*Proxy, er
 			// Management is plain HTTP in-cluster, so allow the proxy token to
 			// ride a non-TLS gRPC connection.
 			"NB_PROXY_ALLOW_INSECURE": "true",
+			// Keep the embedded overlay peer eager (see the client env note):
+			// lazy connections leave peers unconnected until traffic, which
+			// starves the harness's WaitProxyPeer.
+			"NB_LAZY_CONN": "false",
 			// The combined server multiplexes the relay over WebSocket on :8080
 			// (no QUIC listener). The proxy's embedded relay client defaults to
 			// QUIC, which fails here and flaps the relay link, churning the

go.mod

@@ -113,7 +113,7 @@ require (
 	github.com/ti-mo/conntrack v0.5.1
 	github.com/ti-mo/netfilter v0.5.2
 	github.com/vmihailenco/msgpack/v5 v5.4.1
-	github.com/wailsapp/wails/v3 v3.0.0-alpha2.111
+	github.com/wailsapp/wails/v3 v3.0.0-alpha2.117
 	github.com/yusufpapurcu/wmi v1.2.4
 	github.com/zcalusic/sysinfo v1.1.3
 	go.opentelemetry.io/contrib/instrumentation/google.golang.org/grpc/otelgrpc v0.67.0
@@ -303,7 +303,6 @@ require (
 	github.com/tklauser/numcpus v0.10.0 // indirect
 	github.com/vishvananda/netns v0.0.5 // indirect
 	github.com/vmihailenco/tagparser/v2 v2.0.0 // indirect
-	github.com/wailsapp/wails/webview2 v1.0.27 // indirect
 	github.com/wlynxg/anet v0.0.5 // indirect
 	github.com/x448/float16 v0.8.4 // indirect
 	github.com/zeebo/blake3 v0.2.3 // indirect
@@ -335,7 +334,7 @@ replace github.com/cloudflare/circl => codeberg.org/cunicu/circl v0.0.0-20230801
 
 replace github.com/pion/ice/v4 => github.com/netbirdio/ice/v4 v4.0.0-20250908184934-6202be846b51
 
-replace github.com/dexidp/dex => github.com/netbirdio/dex v0.244.1-0.20260512110716-8d70ad8647c1
+replace github.com/dexidp/dex => github.com/netbirdio/dex v0.244.1-0.20260716205454-a163de3129e5
 
 replace github.com/dexidp/dex/api/v2 => github.com/netbirdio/dex/api/v2 v2.0.0-20260512110716-8d70ad8647c1
 

go.sum

@@ -476,8 +476,8 @@ github.com/morikuni/aec v1.0.0 h1:nP9CBfwrvYnBRgY6qfDQkygYDmYwOilePFkwzv4dU8A=
 github.com/morikuni/aec v1.0.0/go.mod h1:BbKIizmSmc5MMPqRYbxO4ZU0S0+P200+tUnFx7PXmsc=
 github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822 h1:C3w9PqII01/Oq1c1nUAm88MOHcQC9l5mIlSMApZMrHA=
 github.com/munnerz/goautoneg v0.0.0-20191010083416-a7dc8b61c822/go.mod h1:+n7T8mK8HuQTcFwEeznm/DIxMOiR9yIdICNftLE1DvQ=
-github.com/netbirdio/dex v0.244.1-0.20260512110716-8d70ad8647c1 h1:4TaYr9O4xX0D2kszeOLclTiCbA3eHq3xWV+9ILJbIYs=
-github.com/netbirdio/dex v0.244.1-0.20260512110716-8d70ad8647c1/go.mod h1:IHH+H8vK2GfqtIt5u/5OdPh18yk0oDHuj2vz5+Goetg=
+github.com/netbirdio/dex v0.244.1-0.20260716205454-a163de3129e5 h1:3PwQv8aR46qN2u16+Dv6udnH3sbVKX5KrGwF35CKSI0=
+github.com/netbirdio/dex v0.244.1-0.20260716205454-a163de3129e5/go.mod h1:IHH+H8vK2GfqtIt5u/5OdPh18yk0oDHuj2vz5+Goetg=
 github.com/netbirdio/dex/api/v2 v2.0.0-20260512110716-8d70ad8647c1 h1:neE7z+FPUkldl3faK/Jt+hJK2L+1XfQ1W33TQhU9m88=
 github.com/netbirdio/dex/api/v2 v2.0.0-20260512110716-8d70ad8647c1/go.mod h1:awuTyT29CYALpEyET0S307EgNlPWrc7fFKRAyhsO45M=
 github.com/netbirdio/easyjson v0.9.0 h1:6Nw2lghSVuy8RSkAYDhDv1thBVEmfVbKZnV7T7Z6Aus=
@@ -660,10 +660,8 @@ github.com/vmihailenco/msgpack/v5 v5.4.1 h1:cQriyiUvjTwOHg8QZaPihLWeRAAVoCpE00IU
 github.com/vmihailenco/msgpack/v5 v5.4.1/go.mod h1:GaZTsDaehaPpQVyxrf5mtQlH+pc21PIudVV/E3rRQok=
 github.com/vmihailenco/tagparser/v2 v2.0.0 h1:y09buUbR+b5aycVFQs/g70pqKVZNBmxwAhO7/IwNM9g=
 github.com/vmihailenco/tagparser/v2 v2.0.0/go.mod h1:Wri+At7QHww0WTrCBeu4J6bNtoV6mEfg5OIWRZA9qds=
-github.com/wailsapp/wails/v3 v3.0.0-alpha2.111 h1:MKx1nOnhnDuEGrRBmtxLOJq1NERwailu2cI4BvzWhi4=
-github.com/wailsapp/wails/v3 v3.0.0-alpha2.111/go.mod h1:wrdvmyeCsB/K3YqJDoH8E3MwcN8NXAMnEFaDTW46w60=
-github.com/wailsapp/wails/webview2 v1.0.27 h1:wjgAi/I8BBZ7kUGU8um3XF3ILEfzr96Q2Q1G4GPjMns=
-github.com/wailsapp/wails/webview2 v1.0.27/go.mod h1:zdM4jcO1IaC61RiJL5F1BzgoqBHFIdacz8gPr5exr0o=
+github.com/wailsapp/wails/v3 v3.0.0-alpha2.117 h1:udyjqPG3AIgkod5QDR/WblCkpV8R86BFPSrsWxSyt5Y=
+github.com/wailsapp/wails/v3 v3.0.0-alpha2.117/go.mod h1:74WH2FScMsgucZvHHvv7eOefDXCm/CjuIxqhhZgPhKg=
 github.com/wlynxg/anet v0.0.5 h1:J3VJGi1gvo0JwZ/P1/Yc/8p63SoW98B5dHkYDmpgvvU=
 github.com/wlynxg/anet v0.0.5/go.mod h1:eay5PRQr7fIVAMbTbchTnO9gG65Hg/uYGdc7mguHxoA=
 github.com/x448/float16 v0.8.4 h1:qLwI1I70+NjRFUR3zs1JPUCgaCXSh3SW62uAKT1mSBM=

idp/dex/config.go

@@ -613,6 +613,10 @@ func (c *YAMLConfig) ToServerConfig(stor storage.Storage, logger *slog.Logger) s
 		cfg.SupportedResponseTypes = c.OAuth2.ResponseTypes
 	}
 
+	if len(c.OAuth2.GrantTypes) > 0 {
+		cfg.AllowedGrantTypes = c.OAuth2.GrantTypes
+	}
+
 	// Apply expiry settings
 	if c.Expiry.IDTokens != "" {
 		if d, err := parseDuration(c.Expiry.IDTokens); err == nil {

idp/dex/provider.go

@@ -21,7 +21,7 @@ import (
 	"github.com/dexidp/dex/server/signer"
 	"github.com/dexidp/dex/storage"
 	"github.com/dexidp/dex/storage/sql"
-	jose "github.com/go-jose/go-jose/v4"
+	"github.com/go-jose/go-jose/v4"
 	"github.com/google/uuid"
 	"github.com/prometheus/client_golang/prometheus"
 	"golang.org/x/crypto/bcrypt"

idp/dex/provider_test.go

@@ -595,3 +595,90 @@ enablePasswordDB: true
 	assert.True(t, cfg.ContinueOnConnectorFailure,
 		"buildDexConfig must set ContinueOnConnectorFailure to true so management starts even if an external IdP is down")
 }
+
+func TestToServerConfig_WiresGrantTypes(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "dex-grants-*")
+	require.NoError(t, err)
+	defer os.RemoveAll(tmpDir)
+
+	stor := openTestStorage(t, tmpDir)
+	defer stor.Close()
+
+	logger := slog.New(slog.NewTextHandler(os.Stderr, &slog.HandlerOptions{Level: slog.LevelError}))
+
+	grants := []string{"authorization_code", "refresh_token"}
+	cfg := &YAMLConfig{Issuer: "http://localhost:5599/oauth2", OAuth2: OAuth2{GrantTypes: grants}}
+	assert.Equal(t, grants, cfg.ToServerConfig(stor, logger).AllowedGrantTypes)
+
+	empty := &YAMLConfig{Issuer: "http://localhost:5599/oauth2"}
+	assert.Empty(t, empty.ToServerConfig(stor, logger).AllowedGrantTypes)
+}
+
+func newDeviceGuardProvider(t *testing.T, grantTypesYAML string) *Provider {
+	t.Helper()
+
+	tmpDir, err := os.MkdirTemp("", "dex-devguard-*")
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = os.RemoveAll(tmpDir) })
+
+	yamlContent := `
+issuer: http://localhost:5599/oauth2
+storage:
+  type: sqlite3
+  config:
+    file: ` + filepath.Join(tmpDir, "dex.db") + `
+web:
+  http: 127.0.0.1:5599
+enablePasswordDB: true
+` + grantTypesYAML
+
+	configPath := filepath.Join(tmpDir, "config.yaml")
+	require.NoError(t, os.WriteFile(configPath, []byte(yamlContent), 0644))
+
+	yamlConfig, err := LoadConfig(configPath)
+	require.NoError(t, err)
+
+	provider, err := NewProviderFromYAML(context.Background(), yamlConfig)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = provider.Stop(context.Background()) })
+	return provider
+}
+
+func TestHandler_BlocksDeviceEndpointsWhenDeviceGrantDisabled(t *testing.T) {
+	provider := newDeviceGuardProvider(t, `
+oauth2:
+  grantTypes:
+    - authorization_code
+    - refresh_token
+`)
+
+	devicePaths := []string{
+		"/oauth2/device",
+		"/oauth2/device/code",
+		"/oauth2/device/token",
+		"/oauth2/device/auth/verify_code",
+		"/oauth2/device/callback",
+	}
+	for _, path := range devicePaths {
+		for _, method := range []string{http.MethodGet, http.MethodPost} {
+			req := httptest.NewRequest(method, path, nil)
+			rec := httptest.NewRecorder()
+			provider.Handler().ServeHTTP(rec, req)
+			assert.Equal(t, http.StatusNotFound, rec.Code, "%s %s must be blocked", method, path)
+		}
+	}
+
+	req := httptest.NewRequest(http.MethodGet, "/oauth2/.well-known/openid-configuration", nil)
+	rec := httptest.NewRecorder()
+	provider.Handler().ServeHTTP(rec, req)
+	assert.Equal(t, http.StatusOK, rec.Code)
+}
+
+func TestHandler_AllowsDeviceEndpointsWhenGrantsDefault(t *testing.T) {
+	provider := newDeviceGuardProvider(t, "")
+
+	req := httptest.NewRequest(http.MethodPost, "/oauth2/device/code", nil)
+	rec := httptest.NewRecorder()
+	provider.Handler().ServeHTTP(rec, req)
+	assert.NotEqual(t, http.StatusNotFound, rec.Code)
+}

management/internals/modules/agentnetwork/catalog/catalog.go

@@ -197,6 +197,12 @@ type JSONMetadataInjection struct {
 	// enforces a 128-char limit per value; oversized values are
 	// truncated rather than failing the request. 0 disables the cap.
 	MaxValueLength int
+	// Sanitize, when true, replaces characters outside the destination's
+	// accepted set with '_' before emitting each value. AWS Bedrock's
+	// X-Amzn-Bedrock-Request-Metadata restricts values to a limited character
+	// class, so unsanitized group display names (e.g. containing spaces) would
+	// make Bedrock reject the request with 400.
+	Sanitize bool
 }
 
 // providers is the canonical list of supported Agent Network providers.
@@ -329,6 +335,18 @@ var providers = []Provider{
 			{ID: "amazon.nova-lite", Label: "Amazon Nova Lite (Bedrock)", InputPer1k: 0.00006, OutputPer1k: 0.00024, ContextWindow: 300000},
 			{ID: "amazon.nova-micro", Label: "Amazon Nova Micro (Bedrock)", InputPer1k: 0.000035, OutputPer1k: 0.00014, ContextWindow: 128000},
 		},
+		// Bedrock accepts a cost-allocation metadata header; stamp the caller's
+		// user + authorizing group so spend can be attributed in AWS Cost
+		// Management. Sanitized because Bedrock restricts the value character set.
+		IdentityInjection: &IdentityInjection{
+			JSONMetadata: &JSONMetadataInjection{
+				Header:         "X-Amzn-Bedrock-Request-Metadata",
+				UserKey:        "user",
+				GroupsKey:      "group",
+				MaxValueLength: 256,
+				Sanitize:       true,
+			},
+		},
 	},
 	{
 		ID:                 "vertex_ai_api",
@@ -402,6 +420,45 @@ var providers = []Provider{
 			{ID: "mistral-embed", Label: "Mistral Embed", InputPer1k: 0.0001, OutputPer1k: 0, ContextWindow: 8192},
 		},
 	},
+	{
+		ID:                 "kimi_api",
+		Kind:               KindProvider,
+		Name:               "Kimi (Moonshot AI) API",
+		Description:        "Kimi K3 / K2 models via the Moonshot AI platform",
+		DefaultHost:        "api.moonshot.ai",
+		AuthHeaderName:     "Authorization",
+		AuthHeaderTemplate: "Bearer ${API_KEY}",
+		DefaultContentType: "application/json",
+		BrandColor:         "#1A1A2E",
+		// ParserID empty on purpose: Moonshot serves two body shapes on
+		// the same host and key, and the proxy's URL sniffer dispatches
+		// both (same pattern as Bifrost). /v1/chat/completions matches
+		// OpenAIParser; the Anthropic-compatible endpoint the official
+		// Claude Code guide uses (/anthropic/v1/messages) contains
+		// "/v1/messages" and matches AnthropicParser. Pinning "openai"
+		// here would misparse the Claude Code path — the primary way
+		// teams consume Kimi for coding today. Both endpoints accept the
+		// same Moonshot key via Authorization: Bearer (Claude Code's
+		// ANTHROPIC_AUTH_TOKEN rides that header too).
+		//
+		// api.moonshot.ai is the international platform; mainland-China
+		// accounts live on api.moonshot.cn with separate billing —
+		// operators there override the host on the provider record. The
+		// kimi.com subscription coding endpoint (api.kimi.com/coding,
+		// model id "k3") is account-bound seat licensing rather than a
+		// meterable platform key, so it's deliberately not the default.
+		ParserID: "",
+		// Pricing per Moonshot's platform rates at K3 launch (July 2026):
+		// K3 $3/$15 per MTok with $0.30 cached input, flat across the
+		// 1M-token window; K2 Thinking $0.60/$2.50 with $0.15 cache hits.
+		// Kimi K3 has a single always-on-reasoning SKU — no mini/turbo
+		// variants at launch. The consumer app's "K3 Swarm Max" mode is
+		// not an API SKU, so it doesn't appear here.
+		Models: []Model{
+			{ID: "kimi-k3", Label: "Kimi K3", InputPer1k: 0.003, OutputPer1k: 0.015, ContextWindow: 1000000},
+			{ID: "kimi-k2-thinking", Label: "Kimi K2 Thinking", InputPer1k: 0.0006, OutputPer1k: 0.0025, ContextWindow: 262144},
+		},
+	},
 	{
 		ID:                 "litellm_proxy",
 		Kind:               KindGateway,

management/internals/modules/agentnetwork/synthesizer.go

@@ -540,6 +540,7 @@ type identityInjectJSONMetadata struct {
 	UserKey        string `json:"user_key,omitempty"`
 	GroupsKey      string `json:"groups_key,omitempty"`
 	MaxValueLength int    `json:"max_value_length,omitempty"`
+	Sanitize       bool   `json:"sanitize,omitempty"`
 }
 
 // buildIdentityInjectConfigJSON walks the enabled providers and emits
@@ -583,9 +584,11 @@ func buildIdentityInjectConfigJSON(providers []*types.Provider, groupIndex map[s
 func buildIdentityInjectRule(p *types.Provider, entry catalog.Provider) (identityInjectProvider, bool) {
 	rule := identityInjectProvider{ProviderID: p.ID}
 	// Identity-stamping shape (one of HeaderPair / JSONMetadata). Skip the
-	// shape silently when the catalog entry doesn't declare one — extras
-	// can still apply, see below.
-	if entry.IdentityInjection != nil {
+	// shape silently when the catalog entry doesn't declare one, or when the
+	// operator disabled metadata for this provider — extras can still apply,
+	// see below. MetadataDisabled suppresses only the identity dimensions
+	// (user + authorizing group), not the catalog's routing ExtraHeaders.
+	if !p.MetadataDisabled && entry.IdentityInjection != nil {
 		switch {
 		case entry.IdentityInjection.HeaderPair != nil:
 			rule.HeaderPair = buildIdentityHeaderPair(p, entry.IdentityInjection.HeaderPair)
@@ -651,6 +654,7 @@ func buildIdentityJSONMetadata(p *types.Provider, jm *catalog.JSONMetadataInject
 		UserKey:        userKey,
 		GroupsKey:      groupsKey,
 		MaxValueLength: jm.MaxValueLength,
+		Sanitize:       jm.Sanitize,
 	}
 }
 

management/internals/modules/agentnetwork/synthesizer_test.go

@@ -698,6 +698,94 @@ func TestSynthesizeServices_IdentityInject_Portkey_NotCustomizable(t *testing.T)
 		"same fixed-schema guarantee for the groups dimension")
 }
 
+// TestSynthesizeServices_IdentityInject_Bedrock pins Bedrock's cost-allocation
+// metadata: a JSONMetadata shape emitting X-Amzn-Bedrock-Request-Metadata with
+// the reserved user/group keys, sanitized to Bedrock's accepted charset.
+func TestSynthesizeServices_IdentityInject_Bedrock(t *testing.T) {
+	ctx := context.Background()
+	ctrl := gomock.NewController(t)
+	defer ctrl.Finish()
+	mockStore := store.NewMockStore(ctrl)
+
+	br := newSynthTestProvider()
+	br.ID = "prov-bedrock"
+	br.ProviderID = "bedrock_api"
+	br.UpstreamURL = "https://bedrock-runtime.us-east-1.amazonaws.com"
+	br.APIKey = "bedrock-bearer"
+	br.CreatedAt = time.Date(2026, 4, 2, 0, 0, 0, 0, time.UTC)
+
+	policy := newSynthTestPolicy(br.ID, "grp-eng", "")
+	policy.ID = "pol-bedrock"
+
+	expectSynthBaseInputs(mockStore, ctx, newSynthTestSettings(),
+		[]*types.Provider{br},
+		[]*types.Policy{policy},
+		[]*types.Guardrail{})
+
+	services, err := SynthesizeServices(ctx, mockStore, testAccountID)
+	require.NoError(t, err)
+	require.Len(t, services, 1)
+
+	var injectCfg identityInjectConfig
+	for _, m := range services[0].Targets[0].Options.Middlewares {
+		if m.ID == middlewareIDLLMIdentityInject {
+			require.NoError(t, json.Unmarshal(m.ConfigJSON, &injectCfg))
+			break
+		}
+	}
+	require.Len(t, injectCfg.Providers, 1)
+	entry := injectCfg.Providers[0]
+	require.NotNil(t, entry.JSONMetadata, "Bedrock uses the JSONMetadata shape for cost-allocation metadata")
+	assert.Nil(t, entry.HeaderPair, "shapes are mutually exclusive")
+	assert.Equal(t, "X-Amzn-Bedrock-Request-Metadata", entry.JSONMetadata.Header,
+		"the caller identity lands in Bedrock's cost-allocation metadata header")
+	assert.Equal(t, "user", entry.JSONMetadata.UserKey)
+	assert.Equal(t, "group", entry.JSONMetadata.GroupsKey)
+	assert.True(t, entry.JSONMetadata.Sanitize,
+		"Bedrock restricts the metadata value charset, so values must be sanitized")
+}
+
+// TestSynthesizeServices_MetadataDisabled_SuppressesInjection verifies the
+// per-provider opt-out: a provider with MetadataDisabled set emits no
+// identity-inject entry (Bedrock has no catalog ExtraHeaders, so the whole
+// entry is dropped).
+func TestSynthesizeServices_MetadataDisabled_SuppressesInjection(t *testing.T) {
+	ctx := context.Background()
+	ctrl := gomock.NewController(t)
+	defer ctrl.Finish()
+	mockStore := store.NewMockStore(ctrl)
+
+	br := newSynthTestProvider()
+	br.ID = "prov-bedrock"
+	br.ProviderID = "bedrock_api"
+	br.UpstreamURL = "https://bedrock-runtime.us-east-1.amazonaws.com"
+	br.APIKey = "bedrock-bearer"
+	br.MetadataDisabled = true
+	br.CreatedAt = time.Date(2026, 4, 2, 0, 0, 0, 0, time.UTC)
+
+	policy := newSynthTestPolicy(br.ID, "grp-eng", "")
+	policy.ID = "pol-bedrock"
+
+	expectSynthBaseInputs(mockStore, ctx, newSynthTestSettings(),
+		[]*types.Provider{br},
+		[]*types.Policy{policy},
+		[]*types.Guardrail{})
+
+	services, err := SynthesizeServices(ctx, mockStore, testAccountID)
+	require.NoError(t, err)
+	require.Len(t, services, 1)
+
+	var injectCfg identityInjectConfig
+	for _, m := range services[0].Targets[0].Options.Middlewares {
+		if m.ID == middlewareIDLLMIdentityInject {
+			require.NoError(t, json.Unmarshal(m.ConfigJSON, &injectCfg))
+			break
+		}
+	}
+	assert.Empty(t, injectCfg.Providers,
+		"metadata_disabled must drop the provider's identity-inject entry")
+}
+
 // TestSynthesizeServices_IdentityInject_Vercel pins Vercel AI
 // Gateway's wiring: HeaderPair shape with fixed wire names dictated
 // by Vercel's Custom Reporting API (ai-reporting-user /

management/internals/modules/agentnetwork/types/provider.go

@@ -51,6 +51,12 @@ type Provider struct {
 	// private or self-signed certificate. The synthesiser propagates it into
 	// the router route so the proxy dials that provider's upstream insecurely.
 	SkipTLSVerification bool `gorm:"column:skip_tls_verification"`
+	// MetadataDisabled suppresses identity metadata injection for this provider.
+	// Metadata (the caller's user + authorizing group) is injected by default;
+	// when true the synthesiser omits the provider's identity-inject shape, so no
+	// user/group headers (e.g. Bedrock's X-Amzn-Bedrock-Request-Metadata) are
+	// stamped. Catalog ExtraHeaders (routing config) are unaffected.
+	MetadataDisabled bool `gorm:"column:metadata_disabled"`
 	// SessionPrivateKey + SessionPublicKey are the ed25519 keypair the
 	// synthesised reverse-proxy service uses to sign / verify session
 	// JWTs after a successful OIDC handshake. Generated once on
@@ -137,6 +143,9 @@ func (p *Provider) FromAPIRequest(req *api.AgentNetworkProviderRequest) {
 	if req.SkipTlsVerification != nil {
 		p.SkipTLSVerification = *req.SkipTlsVerification
 	}
+	if req.MetadataDisabled != nil {
+		p.MetadataDisabled = *req.MetadataDisabled
+	}
 	// Identity-header overrides for catalogs flagged Customizable.
 	// nil pointer = "field omitted on the wire" → leave the stored
 	// value untouched (per the openapi description). Empty string is
@@ -170,6 +179,7 @@ func (p *Provider) ToAPIResponse() *api.AgentNetworkProvider {
 		Models:              models,
 		Enabled:             p.Enabled,
 		SkipTlsVerification: p.SkipTLSVerification,
+		MetadataDisabled:    p.MetadataDisabled,
 		CreatedAt:           &created,
 		UpdatedAt:           &updated,
 	}

management/internals/modules/agentnetwork/types/provider_test.go

@@ -42,3 +42,38 @@ func TestProvider_SkipTLSVerification_RoundTrip(t *testing.T) {
 	assert.False(t, p.SkipTLSVerification, "explicit false must clear skip_tls_verification")
 	assert.False(t, p.ToAPIResponse().SkipTlsVerification, "response must reflect the cleared value")
 }
+
+// TestProvider_MetadataDisabled_RoundTrip covers the request→provider→response
+// mapping of metadata_disabled, with the same update semantics: nil preserves,
+// explicit false clears.
+func TestProvider_MetadataDisabled_RoundTrip(t *testing.T) {
+	enable := true
+	disable := false
+
+	base := func() *api.AgentNetworkProviderRequest {
+		return &api.AgentNetworkProviderRequest{
+			ProviderId:  "bedrock_api",
+			Name:        "bedrock",
+			UpstreamUrl: "https://bedrock-runtime.us-east-1.amazonaws.com",
+		}
+	}
+
+	p := NewProvider("acc-1")
+
+	req := base()
+	req.MetadataDisabled = &enable
+	p.FromAPIRequest(req)
+	assert.True(t, p.MetadataDisabled, "create with metadata_disabled=true must set the field")
+	assert.True(t, p.ToAPIResponse().MetadataDisabled, "response must surface metadata_disabled")
+
+	// Omitting the field on update leaves the stored value untouched.
+	p.FromAPIRequest(base())
+	assert.True(t, p.MetadataDisabled, "omitting metadata_disabled on update must preserve it")
+
+	// Explicit false clears it (re-enables metadata).
+	req = base()
+	req.MetadataDisabled = &disable
+	p.FromAPIRequest(req)
+	assert.False(t, p.MetadataDisabled, "explicit false must clear metadata_disabled")
+	assert.False(t, p.ToAPIResponse().MetadataDisabled, "response must reflect the cleared value")
+}

management/internals/modules/peers/ephemeral/manager/ephemeral.go

@@ -215,7 +215,7 @@ func (e *EphemeralManager) cleanup(ctx context.Context) {
 	}
 
 	for accountID, peerIDs := range peerIDsPerAccount {
-		log.WithContext(ctx).Tracef("cleanup: deleting %d ephemeral peers for account %s", len(peerIDs), accountID)
+		log.WithContext(ctx).Debugf("cleanup: deleting %d ephemeral peers for account %s: %s", len(peerIDs), accountID, peerIDs)
 		err := e.peersManager.DeletePeers(ctx, accountID, peerIDs, activity.SystemInitiator, true)
 		if err != nil {
 			log.WithContext(ctx).Errorf("failed to delete ephemeral peers: %s", err)

management/internals/modules/peers/manager.go

@@ -184,6 +184,8 @@ func (m *managerImpl) DeletePeers(ctx context.Context, accountID string, peerIDs
 				return err
 			}
 
+			log.WithContext(ctx).Debugf("DeletePeers: deleted peer %s", peerID)
+
 			if !(peer.ProxyMeta.Embedded || peer.Meta.KernelVersion == "wasm") {
 				eventsToStore = append(eventsToStore, func() {
 					m.accountManager.StoreEvent(ctx, userID, peer.ID, accountID, activity.PeerRemovedByUser, peer.EventMeta(dnsDomain))
@@ -224,30 +226,6 @@ func (m *managerImpl) CreateProxyPeer(ctx context.Context, accountID string, pee
 		return nil
 	}
 
-	// Dedupe stale embedded peer records for the same (account, cluster).
-	// The proxy generates a fresh WireGuard keypair on every startup
-	// (proxy/internal/roundtrip/netbird.go), so without this sweep the
-	// prior embedded peer would linger forever — holding its CGNAT IP
-	// allocation, polluting other peers' rosters, and (most visibly)
-	// leaving the synth DNS pointing at the dead address. The
-	// (account, cluster) tuple identifies "the embedded peer for this
-	// proxy instance at this cluster"; any record matching that tuple
-	// with a different pubkey is by definition stale and must go.
-	staleIDs, err := m.findStaleEmbeddedProxyPeers(ctx, accountID, cluster, peerKey)
-	if err != nil {
-		return fmt.Errorf("scan for stale embedded proxy peers: %w", err)
-	}
-	if len(staleIDs) > 0 {
-		// userID="" + checkConnected=false: the deletion is initiated
-		// by management itself on behalf of the freshly-registering
-		// proxy, not by an end user; the stale peer may still be
-		// marked Connected from its prior session, but its session is
-		// dead by definition (its key no longer exists).
-		if err := m.DeletePeers(ctx, accountID, staleIDs, "", false); err != nil {
-			return fmt.Errorf("delete stale embedded proxy peers %v: %w", staleIDs, err)
-		}
-	}
-
 	name := fmt.Sprintf("proxy-%s", xid.New().String())
 	newPeer := &peer.Peer{
 		Ephemeral: true,
@@ -273,29 +251,3 @@ func (m *managerImpl) CreateProxyPeer(ctx context.Context, accountID string, pee
 
 	return nil
 }
-
-// findStaleEmbeddedProxyPeers returns the peer IDs of embedded proxy peer
-// records in accountID that target the same cluster but carry a different
-// WireGuard pubkey than the freshly-registering one. Used by CreateProxyPeer
-// to garbage-collect stale records left behind when the proxy restarts with a
-// regenerated keypair.
-func (m *managerImpl) findStaleEmbeddedProxyPeers(ctx context.Context, accountID, cluster, newKey string) ([]string, error) {
-	account, err := m.store.GetAccount(ctx, accountID)
-	if err != nil {
-		return nil, err
-	}
-	var stale []string
-	for _, p := range account.Peers {
-		if p == nil || !p.ProxyMeta.Embedded {
-			continue
-		}
-		if p.ProxyMeta.Cluster != cluster {
-			continue
-		}
-		if p.Key == newKey {
-			continue
-		}
-		stale = append(stale, p.ID)
-	}
-	return stale, nil
-}

management/internals/shared/grpc/token_mgr.go

@@ -161,6 +161,8 @@ func (m *TimeBasedAuthSecretsManager) SetupRefresh(ctx context.Context, accountI
 		m.turnCancelMap[peerID] = turnCancel
 		go m.refreshTURNTokens(ctx, accountID, peerID, turnCancel)
 		log.WithContext(ctx).Debugf("starting TURN refresh for %s", peerID)
+	} else {
+		log.WithContext(ctx).Debugf("no TURN configuration, skipping TURN refresh for %s", peerID)
 	}
 
 	if m.relayCfg != nil {
@@ -168,6 +170,8 @@ func (m *TimeBasedAuthSecretsManager) SetupRefresh(ctx context.Context, accountI
 		m.relayCancelMap[peerID] = relayCancel
 		go m.refreshRelayTokens(ctx, accountID, peerID, relayCancel)
 		log.WithContext(ctx).Tracef("starting relay refresh for %s", peerID)
+	} else {
+		log.WithContext(ctx).Tracef("no relay configuration, skipping relay refresh for %s", peerID)
 	}
 }
 

management/main.go

@@ -1,19 +1,24 @@
 package main
 
 import (
-	"log"
 	"net/http"
 	// nolint:gosec
 	_ "net/http/pprof"
 	"os"
 
+	log "github.com/sirupsen/logrus"
+
 	"github.com/netbirdio/netbird/management/cmd"
 )
 
 func main() {
-	go func() {
-		log.Println(http.ListenAndServe("localhost:6060", nil))
-	}()
+	if pprofAddr := os.Getenv("NB_PPROF_ADDR"); pprofAddr != "" {
+		log.Infof("pprof enabled, listening on: %s", pprofAddr)
+		go func() {
+			log.Println(http.ListenAndServe(pprofAddr, nil))
+		}()
+	}
+
 	if err := cmd.Execute(); err != nil {
 		os.Exit(1)
 	}

management/server/account_test.go

@@ -4269,7 +4269,7 @@ func TestDefaultAccountManager_UpdateAccountSettings_NetworkRangePreserved(t *te
 	}
 
 	// Sanity: an actually different range still triggers reallocation.
-	newRange := netip.MustParsePrefix("100.99.0.0/16")
+	newRange := netip.MustParsePrefix("100.60.0.0/16")
 	_, err = manager.UpdateAccountSettings(ctx, account.Id, userID, &types.Settings{
 		PeerLoginExpirationEnabled: true,
 		PeerLoginExpiration:        types.DefaultPeerLoginExpiration,

management/server/agentnetwork_proxypeer_restart_test.go

@@ -1,199 +0,0 @@
-package server
-
-import (
-	"context"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/management/internals/modules/peers"
-	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
-	agenttypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
-	nbpeer "github.com/netbirdio/netbird/management/server/peer"
-	"github.com/netbirdio/netbird/management/server/permissions"
-	"github.com/netbirdio/netbird/management/server/store"
-	"github.com/netbirdio/netbird/management/server/types"
-)
-
-// TestAgentNetwork_ProxyRestart_PropagatesNewPeerAndDropsStale is the no-mock
-// regression guard for the bug the user reported: restarting the proxy creates
-// a fresh embedded peer with a NEW WireGuard public key (the proxy generates
-// the keypair on every startup at proxy/internal/roundtrip/netbird.go:312).
-// The PRIOR embedded peer record is never deleted on management, so the
-// account accumulates a stale peer holding a stale CGNAT IP. Other peers
-// in the account either keep routing to the dead IP, or — if synth DNS
-// picks the wrong record — never see the new IP at all.
-//
-// What this test exercises (no mocks):
-//   - real SQLite test store
-//   - real DefaultAccountManager, network-map controller, peer-update channels
-//   - real peers.Manager.CreateProxyPeer path (the very method the proxy
-//     invokes over gRPC on every startup)
-//   - real agentnetwork.Manager + synth chain so the client receives a
-//     concrete DNS record that must point at the LATEST proxy peer.
-//
-// Pre-fix expected behavior (red): two embedded peers exist after the
-// "restart"; the synth DNS record points at the stale one; the client
-// receives an update reflecting the new peer but the old one lingers.
-// Post-fix expected behavior (green): exactly one embedded peer exists
-// after restart (with the new key) AND the client's network map carries
-// the synth DNS pointing at that new peer's CGNAT IP.
-func TestAgentNetwork_ProxyRestart_PropagatesNewPeerAndDropsStale(t *testing.T) {
-	am, updateManager, err := createManager(t)
-	require.NoError(t, err, "createManager must succeed")
-	ctx := context.Background()
-
-	const (
-		accountID   = "an-restart-acct"
-		adminUserID = "an-restart-admin"
-		groupAID    = "an-restart-grp-A"
-		clusterAddr = "eu.proxy.netbird.io"
-		clientKey   = "BhRPtynAAYRDy08+q4HTMsos8fs4plTP4NOSh7C1ry8="
-		// Two different proxy pubkeys — the "before" and "after" of a
-		// proxy-process restart with fresh-keypair generation.
-		proxyKey1 = "Aaaaa1aaaaYRDy08+q4HTMsos8fs4plTP4NOSh7C1ry8="
-		proxyKey2 = "Bbbbb2bbbbYRDy08+q4HTMsos8fs4plTP4NOSh7C1ry8="
-	)
-
-	// --- Account scaffold ---
-	account := newAccountWithId(ctx, accountID, adminUserID, "an-restart.test", "", "", false)
-	require.NoError(t, am.Store.SaveAccount(ctx, account))
-
-	clientPeer := &nbpeer.Peer{
-		Key:      clientKey,
-		Name:     "an-restart-client",
-		DNSLabel: "an-restart-client",
-		Meta:     nbpeer.PeerSystemMeta{Hostname: "an-restart-client", GoOS: "linux", WtVersion: "development"},
-	}
-	addedClient, _, _, _, err := am.AddPeer(ctx, "", "", adminUserID, clientPeer, false)
-	require.NoError(t, err, "AddPeer for client must succeed")
-	require.NoError(t, am.MarkPeerConnected(ctx, clientKey, accountID, time.Now().UnixNano(), &types.NetworkMap{}),
-		"MarkPeerConnected for the client peer must succeed (affected-peer fan-out skips disconnected peers)")
-
-	// Place the client in group A so the synth policy reaches it.
-	account, err = am.Store.GetAccount(ctx, accountID)
-	require.NoError(t, err)
-	account.Groups[groupAID] = &types.Group{ID: groupAID, Name: "groupA", Peers: []string{addedClient.ID}}
-	require.NoError(t, am.Store.SaveAccount(ctx, account), "SaveAccount must persist group A")
-
-	// --- Real peers + agent-network managers ---
-	permMgr := permissions.NewManager(am.Store)
-	peersMgr := peers.NewManager(am.Store, permMgr)
-	peersMgr.SetAccountManager(am)
-	peersMgr.SetNetworkMapController(am.networkMapController)
-	agentMgr := agentnetwork.NewManager(am.Store, permMgr, am, nil)
-
-	// Subscribe BEFORE any state-mutating call so we don't lose the update
-	// that contains the synth DNS record.
-	clientCh := updateManager.CreateChannel(ctx, addedClient.ID)
-	t.Cleanup(func() { updateManager.CloseChannel(ctx, addedClient.ID) })
-	drain(clientCh)
-
-	// --- First proxy startup: register peer key K1, then mark it
-	// connected. In production the proxy follows CreateProxyPeer with the
-	// regular sync stream which lands on MarkPeerConnected; the synth DNS
-	// path filters out peers that aren't Connected (types/account.go:323),
-	// so without this step no DNS record would be emitted.
-	require.NoError(t, peersMgr.CreateProxyPeer(ctx, accountID, proxyKey1, clusterAddr),
-		"first CreateProxyPeer (proxy startup) must succeed")
-
-	peer1ID, err := am.Store.GetPeerIDByKey(ctx, store.LockingStrengthNone, proxyKey1)
-	require.NoError(t, err, "proxy peer for K1 must be persisted after CreateProxyPeer")
-	require.NotEmpty(t, peer1ID)
-
-	require.NoError(t, am.MarkPeerConnected(ctx, proxyKey1, accountID, time.Now().UnixNano(), &types.NetworkMap{}),
-		"MarkPeerConnected for K1 must succeed")
-
-	account, err = am.Store.GetAccount(ctx, accountID)
-	require.NoError(t, err)
-	proxyIP1 := account.Peers[peer1ID].IP.String()
-	require.NotEmpty(t, proxyIP1, "K1 must have an assigned overlay IP")
-
-	// --- Provider + policy. CreateProvider / CreatePolicy trigger the
-	// agentnetwork reconcile which runs UpdateAccountPeers; the resulting
-	// NetworkMap delivered to the client carries the synth DNS record
-	// pointing at K1's IP. ---
-	provider, err := agentMgr.CreateProvider(ctx, adminUserID, &agenttypes.Provider{
-		AccountID:   accountID,
-		ProviderID:  "openai_api",
-		Name:        "openai-test",
-		UpstreamURL: "https://api.openai.com",
-		APIKey:      "sk-test-key",
-		Enabled:     true,
-		Models:      []agenttypes.ProviderModel{{ID: "gpt-5.4"}},
-	}, clusterAddr)
-	require.NoError(t, err, "CreateProvider must succeed")
-
-	_, err = agentMgr.CreatePolicy(ctx, adminUserID, &agenttypes.Policy{
-		AccountID:              accountID,
-		Name:                   "p1",
-		Enabled:                true,
-		SourceGroups:           []string{groupAID},
-		DestinationProviderIDs: []string{provider.ID},
-	})
-	require.NoError(t, err, "CreatePolicy must succeed")
-
-	settings, err := am.Store.GetAgentNetworkSettings(ctx, store.LockingStrengthNone, accountID)
-	require.NoError(t, err)
-	fqdn := settings.Endpoint()
-
-	rdata1 := awaitZoneRData(clientCh, clusterAddr, fqdn, true)
-	require.Equal(t, proxyIP1, rdata1,
-		"client must receive a synth DNS record pointing at K1's overlay IP after the synth path runs")
-	drain(clientCh)
-
-	// --- Proxy restart: NEW keypair K2, same account, same cluster ---
-	require.NoError(t, peersMgr.CreateProxyPeer(ctx, accountID, proxyKey2, clusterAddr),
-		"second CreateProxyPeer (proxy restart with fresh keypair) must succeed")
-
-	peer2ID, err := am.Store.GetPeerIDByKey(ctx, store.LockingStrengthNone, proxyKey2)
-	require.NoError(t, err, "proxy peer for K2 must be persisted after restart")
-	require.NotEmpty(t, peer2ID)
-
-	require.NoError(t, am.MarkPeerConnected(ctx, proxyKey2, accountID, time.Now().UnixNano(), &types.NetworkMap{}),
-		"MarkPeerConnected for K2 must succeed")
-
-	// In production the agent's sync stream pulls a fresh NetworkMap as
-	// part of its normal reconcile cadence; in this isolated test
-	// MarkPeerConnected's affected-peer fan-out can race the channel-side
-	// buffer in a way that swallows the synth-DNS-bearing update before
-	// our await reads it. Trigger an explicit account-wide fan-out so the
-	// assertion below tests what production actually delivers, not the
-	// in-test buffer race.
-	am.UpdateAccountPeers(ctx, accountID, types.UpdateReason{Resource: types.UpdateResourcePeer, Operation: types.UpdateOperationUpdate})
-
-	account, err = am.Store.GetAccount(ctx, accountID)
-	require.NoError(t, err)
-	proxyIP2 := account.Peers[peer2ID].IP.String()
-	require.NotEmpty(t, proxyIP2, "K2 must have an assigned overlay IP")
-	require.NotEqual(t, proxyIP1, proxyIP2, "K2 must get a different overlay IP than K1 (sanity)")
-
-	// CRITICAL ASSERTION 1: K1 must no longer be in the store. The SqlStore
-	// returns ("", nil) for a missing key rather than NotFound, so assert
-	// on the returned ID being empty.
-	staleID, err := am.Store.GetPeerIDByKey(ctx, store.LockingStrengthNone, proxyKey1)
-	require.NoError(t, err, "GetPeerIDByKey for a missing peer must not error")
-	assert.Empty(t, staleID,
-		"stale embedded proxy peer K1 must be removed when a new embedded peer registers for the same (account, cluster); pre-fix this assertion fails because management never cleans up the prior peer record")
-
-	// CRITICAL ASSERTION 2: exactly one embedded proxy peer remains, and it
-	// is K2.
-	account, err = am.Store.GetAccount(ctx, accountID)
-	require.NoError(t, err)
-	embeddedKeys := []string{}
-	for _, p := range account.Peers {
-		if p.ProxyMeta.Embedded {
-			embeddedKeys = append(embeddedKeys, p.Key)
-		}
-	}
-	assert.Equal(t, []string{proxyKey2}, embeddedKeys,
-		"after a proxy restart exactly one embedded proxy peer should remain — the one with the new key K2")
-
-	// CRITICAL ASSERTION 3: the synth DNS record the client receives now
-	// points at K2's IP, not K1's.
-	rdata2 := awaitZoneRData(clientCh, clusterAddr, fqdn, true)
-	assert.Equal(t, proxyIP2, rdata2,
-		"after proxy restart, the client's synth DNS record must point at the NEW embedded peer's IP, not the stale K1 IP")
-}

management/server/http/handlers/accounts/accounts_handler.go

@@ -289,6 +289,18 @@ func (h *handler) updateAccountRequestSettings(req api.PutApiAccountsAccountIdJS
 	if req.Settings.AgentNetworkOnly != nil {
 		returnSettings.AgentNetworkOnly = *req.Settings.AgentNetworkOnly
 	}
+	if req.Settings.DashboardFeatures != nil {
+		returnSettings.DashboardFeatures = &types.DashboardFeatures{
+			AgentNetwork: req.Settings.DashboardFeatures.AgentNetwork,
+		}
+	}
+
+	if returnSettings.AgentNetworkOnly &&
+		(returnSettings.DashboardFeatures == nil ||
+			returnSettings.DashboardFeatures.AgentNetwork == nil ||
+			!*returnSettings.DashboardFeatures.AgentNetwork) {
+		return nil, status.Errorf(status.InvalidArgument, "agent network only mode requires dashboard_features.agent_network to be enabled")
+	}
 
 	return returnSettings, nil
 }
@@ -434,6 +446,11 @@ func toAccountResponse(accountID string, settings *types.Settings, meta *types.A
 		networkRangeV6Str := settings.NetworkRangeV6.String()
 		apiSettings.NetworkRangeV6 = &networkRangeV6Str
 	}
+	if settings.DashboardFeatures != nil {
+		apiSettings.DashboardFeatures = &api.AccountDashboardFeatures{
+			AgentNetwork: settings.DashboardFeatures.AgentNetwork,
+		}
+	}
 
 	apiOnboarding := api.AccountOnboarding{
 		OnboardingFlowPending: onboarding.OnboardingFlowPending,

management/server/http/handlers/accounts/accounts_handler_test.go

@@ -288,7 +288,7 @@ func TestAccounts_AccountsHandler(t *testing.T) {
 			expectedBody:   true,
 			requestType:    http.MethodPut,
 			requestPath:    "/api/accounts/" + accountID,
-			requestBody:    bytes.NewBufferString("{\"settings\": {\"peer_login_expiration\": 15552000,\"peer_login_expiration_enabled\": true,\"agent_network_only\": true},\"onboarding\": {\"onboarding_flow_pending\": true,\"signup_form_pending\": true}}"),
+			requestBody:    bytes.NewBufferString("{\"settings\": {\"peer_login_expiration\": 15552000,\"peer_login_expiration_enabled\": true,\"agent_network_only\": true,\"dashboard_features\": {\"agent_network\": true}},\"onboarding\": {\"onboarding_flow_pending\": true,\"signup_form_pending\": true}}"),
 			expectedStatus: http.StatusOK,
 			expectedSettings: api.AccountSettings{
 				PeerLoginExpiration:             15552000,
@@ -305,9 +305,53 @@ func TestAccounts_AccountsHandler(t *testing.T) {
 				AutoUpdateVersion:               sr(""),
 				MetricsPushEnabled:              br(false),
 				AgentNetworkOnly:                br(true),
-				EmbeddedIdpEnabled:              br(false),
-				LocalAuthDisabled:               br(false),
-				LocalMfaEnabled:                 br(false),
+				DashboardFeatures: &api.AccountDashboardFeatures{
+					AgentNetwork: br(true),
+				},
+				EmbeddedIdpEnabled: br(false),
+				LocalAuthDisabled:  br(false),
+				LocalMfaEnabled:    br(false),
+			},
+			expectedArray: false,
+			expectedID:    accountID,
+		},
+		{
+			name:           "PutAccount fails enabling agent_network_only without dashboard_features",
+			expectedBody:   true,
+			requestType:    http.MethodPut,
+			requestPath:    "/api/accounts/" + accountID,
+			requestBody:    bytes.NewBufferString("{\"settings\": {\"peer_login_expiration\": 15552000,\"peer_login_expiration_enabled\": true,\"agent_network_only\": true},\"onboarding\": {\"onboarding_flow_pending\": true,\"signup_form_pending\": true}}"),
+			expectedStatus: http.StatusUnprocessableEntity,
+			expectedArray:  false,
+		},
+		{
+			name:           "PutAccount OK setting dashboard_features agent_network",
+			expectedBody:   true,
+			requestType:    http.MethodPut,
+			requestPath:    "/api/accounts/" + accountID,
+			requestBody:    bytes.NewBufferString("{\"settings\": {\"peer_login_expiration\": 15552000,\"peer_login_expiration_enabled\": true,\"dashboard_features\": {\"agent_network\": true}},\"onboarding\": {\"onboarding_flow_pending\": true,\"signup_form_pending\": true}}"),
+			expectedStatus: http.StatusOK,
+			expectedSettings: api.AccountSettings{
+				PeerLoginExpiration:             15552000,
+				PeerLoginExpirationEnabled:      true,
+				GroupsPropagationEnabled:        br(false),
+				JwtGroupsClaimName:              sr(""),
+				JwtGroupsEnabled:                br(false),
+				JwtAllowGroups:                  &[]string{},
+				RegularUsersViewBlocked:         false,
+				RoutingPeerDnsResolutionEnabled: br(false),
+				LazyConnectionEnabled:           br(false),
+				DnsDomain:                       sr(""),
+				AutoUpdateAlways:                br(false),
+				AutoUpdateVersion:               sr(""),
+				MetricsPushEnabled:              br(false),
+				AgentNetworkOnly:                br(false),
+				DashboardFeatures: &api.AccountDashboardFeatures{
+					AgentNetwork: br(true),
+				},
+				EmbeddedIdpEnabled: br(false),
+				LocalAuthDisabled:  br(false),
+				LocalMfaEnabled:    br(false),
 			},
 			expectedArray: false,
 			expectedID:    accountID,

management/server/idp/embedded.go

@@ -76,6 +76,9 @@ type EmbeddedIdPConfig struct {
 	DashboardPostLogoutRedirectURIs []string
 	// StaticConnectors are additional connectors to seed during initialization
 	StaticConnectors []dex.Connector
+	// GrantTypes restricts allowed OAuth2 grants; empty means all (Dex default). Omit the
+	// device_code grant to disable the device flow; keep authorization_code and refresh_token.
+	GrantTypes []string
 }
 
 // EmbeddedStorageConfig holds storage configuration for the embedded IdP.
@@ -175,6 +178,7 @@ func (c *EmbeddedIdPConfig) ToYAMLConfig() (*dex.YAMLConfig, error) {
 		},
 		OAuth2: dex.OAuth2{
 			SkipApprovalScreen: true,
+			GrantTypes:         c.GrantTypes,
 		},
 		Frontend: dex.Frontend{
 			Issuer: "NetBird",

management/server/peer.go

@@ -106,11 +106,13 @@ func (am *DefaultAccountManager) MarkPeerConnected(ctx context.Context, peerPubK
 	}
 	if !updated {
 		am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusConnect, telemetry.PeerStatusStale)
-		log.WithContext(ctx).Tracef("peer %s already has a newer session in store, skipping connect", peer.ID)
+		log.WithContext(ctx).Debugf("peer %s already has a newer session in store, skipping connect", peer.ID)
 		return nil
 	}
 	am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusConnect, telemetry.PeerStatusApplied)
 
+	log.WithContext(ctx).Debugf("mark peer %s connected", peer.ID)
+
 	if err = am.schedulePeerExpirations(ctx, accountID, peer); err != nil {
 		return err
 	}
@@ -180,12 +182,14 @@ func (am *DefaultAccountManager) MarkPeerDisconnected(ctx context.Context, peerP
 	}
 	if !updated {
 		am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusDisconnect, telemetry.PeerStatusStale)
-		log.WithContext(ctx).Tracef("peer %s session token mismatch on disconnect (token=%d), skipping",
+		log.WithContext(ctx).Debugf("peer %s session token mismatch on disconnect (token=%d), skipping",
 			peer.ID, sessionStartedAt)
 		return nil
 	}
 	am.metrics.AccountManagerMetrics().CountPeerStatusUpdate(telemetry.PeerStatusDisconnect, telemetry.PeerStatusApplied)
 
+	log.WithContext(ctx).Debugf("mark peer %s disconnected", peer.ID)
+
 	// Symmetric with MarkPeerConnected: when an embedded proxy peer goes
 	// offline, refresh the peers that had synthesized records pointing at
 	// it so they pull the stale entries instead of waiting out TTL.

management/server/store/sql_store.go

@@ -1606,6 +1606,8 @@ func (s *SqlStore) getAccount(ctx context.Context, accountID string) (*types.Acc
 			settings_routing_peer_dns_resolution_enabled, settings_dns_domain, settings_network_range,
 			settings_network_range_v6, settings_ipv6_enabled_groups, settings_lazy_connection_enabled,
 			settings_local_mfa_enabled, settings_metrics_push_enabled, settings_agent_network_only,
+			settings_dashboard_features, settings_auto_update_version, settings_auto_update_always,
+			settings_peer_expose_enabled, settings_peer_expose_groups,
 			-- Embedded ExtraSettings
 			settings_extra_peer_approval_enabled, settings_extra_user_approval_required,
 			settings_extra_integrated_validator, settings_extra_integrated_validator_groups
@@ -1630,6 +1632,11 @@ func (s *SqlStore) getAccount(ctx context.Context, accountID string) (*types.Acc
 		sLocalMFAEnabled                 sql.NullBool
 		sMetricsPushEnabled              sql.NullBool
 		sAgentNetworkOnly                sql.NullBool
+		sDashboardFeatures               sql.NullString
+		autoUpdateVersion                sql.NullString
+		autoUpdateAlways                 sql.NullBool
+		peerExposeEnabled                sql.NullBool
+		peerExposeGroups                 sql.NullString
 		sExtraPeerApprovalEnabled        sql.NullBool
 		sExtraUserApprovalRequired       sql.NullBool
 		sExtraIntegratedValidator        sql.NullString
@@ -1653,6 +1660,8 @@ func (s *SqlStore) getAccount(ctx context.Context, accountID string) (*types.Acc
 		&sRoutingPeerDNSResolutionEnabled, &sDNSDomain, &sNetworkRange,
 		&sNetworkRangeV6, &sIPv6EnabledGroups, &sLazyConnectionEnabled,
 		&sLocalMFAEnabled, &sMetricsPushEnabled, &sAgentNetworkOnly,
+		&sDashboardFeatures, &autoUpdateVersion, &autoUpdateAlways,
+		&peerExposeEnabled, &peerExposeGroups,
 		&sExtraPeerApprovalEnabled, &sExtraUserApprovalRequired,
 		&sExtraIntegratedValidator, &sExtraIntegratedValidatorGroups,
 	)
@@ -1724,6 +1733,11 @@ func (s *SqlStore) getAccount(ctx context.Context, accountID string) (*types.Acc
 	if sAgentNetworkOnly.Valid {
 		account.Settings.AgentNetworkOnly = sAgentNetworkOnly.Bool
 	}
+	if sDashboardFeatures.Valid && sDashboardFeatures.String != "" {
+		if err := json.Unmarshal([]byte(sDashboardFeatures.String), &account.Settings.DashboardFeatures); err != nil {
+			log.WithContext(ctx).Warnf("failed to unmarshal dashboard features for account %s: %v", accountID, err)
+		}
+	}
 	if sJWTAllowGroups.Valid {
 		_ = json.Unmarshal([]byte(sJWTAllowGroups.String), &account.Settings.JWTAllowGroups)
 	}
@@ -1739,6 +1753,18 @@ func (s *SqlStore) getAccount(ctx context.Context, accountID string) (*types.Acc
 	if sIPv6EnabledGroups.Valid {
 		_ = json.Unmarshal([]byte(sIPv6EnabledGroups.String), &account.Settings.IPv6EnabledGroups)
 	}
+	if autoUpdateAlways.Valid {
+		account.Settings.AutoUpdateAlways = autoUpdateAlways.Bool
+	}
+	if autoUpdateVersion.Valid {
+		account.Settings.AutoUpdateVersion = autoUpdateVersion.String
+	}
+	if peerExposeEnabled.Valid {
+		account.Settings.PeerExposeEnabled = peerExposeEnabled.Bool
+	}
+	if peerExposeGroups.Valid {
+		_ = json.Unmarshal([]byte(peerExposeGroups.String), &account.Settings.PeerExposeGroups)
+	}
 
 	if sExtraPeerApprovalEnabled.Valid {
 		account.Settings.Extra.PeerApprovalEnabled = sExtraPeerApprovalEnabled.Bool

management/server/store/sql_store_test.go

@@ -9,6 +9,7 @@ import (
 	"net"
 	"net/netip"
 	"os"
+	"reflect"
 	"runtime"
 	"sort"
 	"sync"
@@ -34,6 +35,7 @@ import (
 	"github.com/netbirdio/netbird/management/server/util"
 	nbroute "github.com/netbirdio/netbird/route"
 	"github.com/netbirdio/netbird/shared/management/status"
+	"github.com/netbirdio/netbird/shared/testing_helpers"
 	"github.com/netbirdio/netbird/util/crypt"
 )
 
@@ -296,6 +298,53 @@ func Test_SaveAccount(t *testing.T) {
 	})
 }
 
+func Test_AccountSettings_SaveAndRetrieve(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("The SQLite store is not properly supported by Windows yet")
+	}
+
+	populateFields := testing_helpers.NewPopulateFields().WithCustomFieldSetter(
+		reflect.PointerTo(reflect.TypeOf(types.ExtraSettings{})), func(this *testing_helpers.PopulateFields, field reflect.Value) (int, error) {
+			es := types.ExtraSettings{}
+			reflectedEs := reflect.ValueOf(&es).Elem()
+			n, err := this.PopulateAll(reflectedEs)
+			if err != nil {
+				return n, err
+			}
+			field.Set(reflectedEs.Addr())
+			return n, nil
+		}).WithCustomFieldSetter(
+		reflect.PointerTo(reflect.TypeOf(types.DashboardFeatures{})), func(this *testing_helpers.PopulateFields, field reflect.Value) (int, error) {
+			t := true
+			df := types.DashboardFeatures{AgentNetwork: &t}
+			reflectedDf := reflect.ValueOf(&df).Elem()
+			field.Set(reflectedDf.Addr())
+			return 1, nil
+		}).WithSkippedTag("gorm", "-")
+
+	runTestForAllEngines(t, "", func(t *testing.T, store Store) {
+		account := newAccountWithId(context.Background(), "account_id", "testuser", "")
+		setupKey, _ := types.GenerateDefaultSetupKey()
+		account.SetupKeys[setupKey.Key] = setupKey
+
+		settings := types.Settings{}
+		numOfExportedFields, err := populateFields.PopulateAll(reflect.ValueOf(&settings).Elem())
+		assert.NoError(t, err)
+		assert.Equal(t, 27, numOfExportedFields)
+		account.Settings = &settings
+
+		err = store.SaveAccount(context.Background(), account)
+		assert.NoError(t, err)
+
+		accountFromDb, err := store.GetAccount(context.Background(), account.Id)
+		assert.NoError(t, err)
+		assert.NotNil(t, accountFromDb)
+		assert.NotNil(t, accountFromDb.Settings)
+
+		assert.True(t, reflect.DeepEqual(&settings, accountFromDb.Settings), "created settings and settings retrieved from the db should match")
+	})
+}
+
 func TestSqlite_DeleteAccount(t *testing.T) {
 	if runtime.GOOS == "windows" {
 		t.Skip("The SQLite store is not properly supported by Windows yet")
@@ -1270,6 +1319,36 @@ func TestSqlStore_SaveAccountPersistsAgentNetworkOnly(t *testing.T) {
 	require.False(t, disabled.Settings.AgentNetworkOnly, "disabling should persist")
 }
 
+func TestSqlStore_SaveAccountPersistsDashboardFeatures(t *testing.T) {
+	store, cleanup, err := NewTestStoreFromSQL(context.Background(), "../testdata/store.sql", t.TempDir())
+	t.Cleanup(cleanup)
+	require.NoError(t, err)
+
+	accountID := "bf1c8084-ba50-4ce7-9439-34653001fc3b"
+	account, err := store.GetAccount(context.Background(), accountID)
+	require.NoError(t, err)
+	require.Nil(t, account.Settings.DashboardFeatures, "dashboard features should default to unset")
+
+	agentNetwork := true
+	account.Settings.DashboardFeatures = &types.DashboardFeatures{AgentNetwork: &agentNetwork}
+	require.NoError(t, store.SaveAccount(context.Background(), account))
+
+	reloaded, err := store.GetAccount(context.Background(), accountID)
+	require.NoError(t, err)
+	require.NotNil(t, reloaded.Settings.DashboardFeatures, "dashboard features should survive a save/load round-trip")
+	require.NotNil(t, reloaded.Settings.DashboardFeatures.AgentNetwork, "agent network flag should be set")
+	require.True(t, *reloaded.Settings.DashboardFeatures.AgentNetwork, "agent network flag should persist as true")
+
+	disabled := false
+	reloaded.Settings.DashboardFeatures = &types.DashboardFeatures{AgentNetwork: &disabled}
+	require.NoError(t, store.SaveAccount(context.Background(), reloaded))
+
+	reloadedDisabled, err := store.GetAccount(context.Background(), accountID)
+	require.NoError(t, err)
+	require.NotNil(t, reloadedDisabled.Settings.DashboardFeatures.AgentNetwork, "agent network flag should remain set")
+	require.False(t, *reloadedDisabled.Settings.DashboardFeatures.AgentNetwork, "explicit false should persist")
+}
+
 func TestSqlStore_GetAccountUsers(t *testing.T) {
 	store, cleanup, err := NewTestStoreFromSQL(context.Background(), "../testdata/extended-store.sql", t.TempDir())
 	t.Cleanup(cleanup)

management/server/types/networkmap_components.go

@@ -7,6 +7,7 @@ import (
 	"slices"
 	"strconv"
 	"strings"
+	"sync"
 	"time"
 
 	"github.com/netbirdio/netbird/client/ssh/auth"
@@ -42,6 +43,14 @@ type NetworkMapComponents struct {
 	PostureFailedPeers map[string]map[string]struct{}
 
 	RouterPeers map[string]*nbpeer.Peer
+
+	routesByPeerOnce sync.Once
+	routesByPeerIdx  map[string][]routeIndexEntry
+}
+
+type routeIndexEntry struct {
+	route    *route.Route
+	viaGroup bool
 }
 
 type AccountSettingsInfo struct {
@@ -530,33 +539,43 @@ func (c *NetworkMapComponents) getRoutingPeerRoutes(peerID string) (enabledRoute
 		disabledRoutes = append(disabledRoutes, r)
 	}
 
-	for _, r := range c.Routes {
-		for _, groupID := range r.PeerGroups {
-			group := c.GetGroupInfo(groupID)
-			if group == nil {
-				continue
-			}
-			for _, id := range group.Peers {
-				if id != peerID {
-					continue
-				}
-
-				newPeerRoute := r.Copy()
-				newPeerRoute.Peer = id
-				newPeerRoute.PeerGroups = nil
-				newPeerRoute.ID = route.ID(string(r.ID) + ":" + id)
-				takeRoute(newPeerRoute)
-				break
-			}
-		}
-		if r.Peer == peerID {
-			takeRoute(r.Copy())
+	for _, entry := range c.routesByPeer()[peerID] {
+		if entry.viaGroup {
+			newPeerRoute := entry.route.Copy()
+			newPeerRoute.PeerGroups = nil
+			newPeerRoute.ID = route.ID(string(entry.route.ID) + ":" + peerID)
+			takeRoute(newPeerRoute)
+			continue
 		}
+		takeRoute(entry.route.Copy())
 	}
 
 	return enabledRoutes, disabledRoutes
 }
 
+func (c *NetworkMapComponents) routesByPeer() map[string][]routeIndexEntry {
+	c.routesByPeerOnce.Do(func() {
+		idx := make(map[string][]routeIndexEntry)
+		for _, r := range c.Routes {
+			for _, groupID := range r.PeerGroups {
+				group := c.GetGroupInfo(groupID)
+				if group == nil {
+					continue
+				}
+				for _, id := range group.Peers {
+					idx[id] = append(idx[id], routeIndexEntry{route: r, viaGroup: true})
+				}
+			}
+			if r.Peer != "" {
+				idx[r.Peer] = append(idx[r.Peer], routeIndexEntry{route: r})
+			}
+		}
+		c.routesByPeerIdx = idx
+	})
+
+	return c.routesByPeerIdx
+}
+
 func (c *NetworkMapComponents) filterRoutesByGroups(routes []*route.Route, groupListMap LookupMap) []*route.Route {
 	var filteredRoutes []*route.Route
 	for _, r := range routes {

management/server/types/settings.go

@@ -80,6 +80,11 @@ type Settings struct {
 	// Set for accounts created via netbird.ai signups; users can disable it later.
 	AgentNetworkOnly bool `gorm:"default:false"`
 
+	// DashboardFeatures holds per-account dashboard section visibility overrides.
+	// It serializes to a single JSON column so new sections can be added without
+	// a schema change.
+	DashboardFeatures *DashboardFeatures `gorm:"serializer:json"`
+
 	// EmbeddedIdpEnabled indicates if the embedded identity provider is enabled.
 	// This is a runtime-only field, not stored in the database.
 	EmbeddedIdpEnabled bool `gorm:"-"`
@@ -126,9 +131,31 @@ func (s *Settings) Copy() *Settings {
 	if s.Extra != nil {
 		settings.Extra = s.Extra.Copy()
 	}
+	if s.DashboardFeatures != nil {
+		settings.DashboardFeatures = s.DashboardFeatures.Copy()
+	}
 	return settings
 }
 
+// DashboardFeatures holds per-account dashboard section visibility overrides.
+// Nil fields are unset and follow the default dashboard behavior; an explicit
+// value forces that section shown or hidden for the account.
+type DashboardFeatures struct {
+	// AgentNetwork, when set, forces the Agent Network menu shown (true) or
+	// hidden (false) regardless of the deployment feature flag.
+	AgentNetwork *bool `json:"agent_network,omitempty"`
+}
+
+// Copy returns a deep copy of the DashboardFeatures struct.
+func (d *DashboardFeatures) Copy() *DashboardFeatures {
+	c := &DashboardFeatures{}
+	if d.AgentNetwork != nil {
+		v := *d.AgentNetwork
+		c.AgentNetwork = &v
+	}
+	return c
+}
+
 type ExtraSettings struct {
 	// PeerApprovalEnabled enables or disables the need for peers bo be approved by an administrator
 	PeerApprovalEnabled bool

proxy/Dockerfile.multistage

@@ -14,6 +14,7 @@ COPY proxy ./proxy
 COPY route ./route
 COPY shared ./shared
 COPY sharedsock ./sharedsock
+COPY trustedproxy ./trustedproxy
 COPY upload-server ./upload-server
 COPY util ./util
 COPY version ./version

proxy/cmd/proxy/cmd/root.go

@@ -18,6 +18,7 @@ import (
 	"github.com/netbirdio/netbird/client/embed"
 	"github.com/netbirdio/netbird/proxy"
 	nbacme "github.com/netbirdio/netbird/proxy/internal/acme"
+	"github.com/netbirdio/netbird/trustedproxy"
 	"github.com/netbirdio/netbird/util"
 )
 
@@ -209,7 +210,7 @@ func runServer(cmd *cobra.Command, args []string) error {
 		return fmt.Errorf("invalid domain value %q: %w", proxyDomain, err)
 	}
 
-	parsedTrustedProxies, err := proxy.ParseTrustedProxies(trustedProxies)
+	parsedTrustedProxies, err := trustedproxy.Parse(trustedProxies)
 	if err != nil {
 		return fmt.Errorf("invalid --trusted-proxies: %w", err)
 	}

proxy/internal/accesslog/logger.go

@@ -16,6 +16,7 @@ import (
 	"github.com/netbirdio/netbird/proxy/auth"
 	"github.com/netbirdio/netbird/proxy/internal/types"
 	"github.com/netbirdio/netbird/shared/management/proto"
+	"github.com/netbirdio/netbird/trustedproxy"
 )
 
 const (
@@ -66,7 +67,7 @@ type denyBucket struct {
 type Logger struct {
 	client         gRPCClient
 	logger         *log.Logger
-	trustedProxies []netip.Prefix
+	trustedProxies *trustedproxy.List
 
 	usageMux    sync.Mutex
 	domainUsage map[string]*domainUsage
@@ -82,7 +83,7 @@ type Logger struct {
 // NewLogger creates a new access log Logger. The trustedProxies parameter
 // configures which upstream proxy IP ranges are trusted for extracting
 // the real client IP from X-Forwarded-For headers.
-func NewLogger(client gRPCClient, logger *log.Logger, trustedProxies []netip.Prefix) *Logger {
+func NewLogger(client gRPCClient, logger *log.Logger, trustedProxies *trustedproxy.List) *Logger {
 	if logger == nil {
 		logger = log.StandardLogger()
 	}

proxy/internal/accesslog/requestip.go

@@ -4,13 +4,13 @@ import (
 	"net/http"
 	"net/netip"
 
-	"github.com/netbirdio/netbird/proxy/internal/proxy"
+	"github.com/netbirdio/netbird/trustedproxy"
 )
 
 // extractSourceIP resolves the real client IP from the request using trusted
 // proxy configuration. When trustedProxies is non-empty and the direct
 // connection is from a trusted source, it walks X-Forwarded-For right-to-left
 // skipping trusted IPs. Otherwise it returns RemoteAddr directly.
-func extractSourceIP(r *http.Request, trustedProxies []netip.Prefix) netip.Addr {
-	return proxy.ResolveClientIP(r.RemoteAddr, r.Header.Get("X-Forwarded-For"), trustedProxies)
+func extractSourceIP(r *http.Request, trustedProxies *trustedproxy.List) netip.Addr {
+	return trustedProxies.ResolveClientIP(r.RemoteAddr, r.Header.Get("X-Forwarded-For"))
 }

proxy/internal/llm/bedrock_model.go

@@ -0,0 +1,38 @@
+package llm
+
+import (
+	"regexp"
+	"strings"
+)
+
+// bedrockRegionPrefixes are the cross-region inference-profile prefixes that
+// front a Bedrock model id (e.g. "eu.anthropic.claude-...").
+var bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
+
+// bedrockVersionSuffix matches the trailing "-vN[:N]" or "-YYYYMMDD-vN[:N]"
+// version/throughput suffix of a Bedrock model id.
+var bedrockVersionSuffix = regexp.MustCompile(`-(\d{8}-)?v\d+(:\d+)?$`)
+
+// NormalizeBedrockModel strips an ARN wrapper, a cross-region inference-profile
+// prefix, and the version/throughput suffix from a Bedrock model id so it
+// matches the catalog/pricing key, e.g.
+// "eu.anthropic.claude-sonnet-4-5-20250929-v1:0" -> "anthropic.claude-sonnet-4-5"
+// and the inference-profile ARN's last segment likewise. It is the single
+// source of truth shared by the request parser (which normalizes the request
+// model from the URL path) and the router (which normalizes the operator's
+// registered Bedrock model ids so both sides compare equal).
+func NormalizeBedrockModel(modelID string) string {
+	m := modelID
+	if strings.HasPrefix(m, "arn:") {
+		if i := strings.LastIndex(m, "/"); i >= 0 {
+			m = m[i+1:]
+		}
+	}
+	for _, p := range bedrockRegionPrefixes {
+		if strings.HasPrefix(m, p) {
+			m = m[len(p):]
+			break
+		}
+	}
+	return bedrockVersionSuffix.ReplaceAllString(m, "")
+}

proxy/internal/llm/bedrock_model_test.go

@@ -0,0 +1,23 @@
+package llm
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestNormalizeBedrockModel(t *testing.T) {
+	cases := map[string]string{
+		"eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
+		"us.anthropic.claude-haiku-4-5":                "anthropic.claude-haiku-4-5",
+		"us.anthropic.claude-opus-4-8-20250101-v1:0":   "anthropic.claude-opus-4-8",
+		"anthropic.claude-sonnet-4-5-20250929-v1:0":    "anthropic.claude-sonnet-4-5",
+		"meta.llama3-3-70b-instruct-v1:0":              "meta.llama3-3-70b-instruct",
+		"amazon.nova-pro-v1:0":                         "amazon.nova-pro",
+		// Inference-profile ARN — model id lives in the last path segment.
+		"arn:aws:bedrock:eu-central-1:123456789012:inference-profile/eu.anthropic.claude-sonnet-4-5-20250929-v1:0": "anthropic.claude-sonnet-4-5",
+	}
+	for in, want := range cases {
+		require.Equal(t, want, NormalizeBedrockModel(in), "normalize %q", in)
+	}
+}

proxy/internal/llm/pricing/defaults_pricing.yaml

@@ -161,6 +161,18 @@ openai:
     input_per_1k: 0.0001
     output_per_1k: 0
 
+  # Kimi / Moonshot AI (kimi_api) — OpenAI-compatible /v1 endpoint. Moonshot
+  # reports cache hits OpenAI-style when present; cached input is 10% of
+  # input for K3 ($0.30 vs $3.00 per MTok) and $0.15/MTok for K2 Thinking.
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cached_input_per_1k: 0.0003
+  kimi-k2-thinking:
+    input_per_1k: 0.0006
+    output_per_1k: 0.0025
+    cached_input_per_1k: 0.00015
+
 anthropic:
   # Claude 4.x family — cache reads ≈10% of input, cache writes ≈125% of input.
   # Pricing source: Anthropic's current published rates per million tokens,
@@ -206,6 +218,24 @@ anthropic:
     cache_read_per_1k: 0.0001
     cache_creation_per_1k: 0.00125
 
+  # Kimi / Moonshot AI (kimi_api) via the Anthropic-compatible endpoint
+  # (/anthropic/v1/messages — the official Claude Code setup). Same rates
+  # as the OpenAI-shape entries above. "kimi-k3[1m]" is the model id some
+  # Claude Code guides set for the 1M-context alias; priced identically so
+  # cost metering doesn't silently skip those requests.
+  kimi-k3:
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+  "kimi-k3[1m]":
+    input_per_1k: 0.003
+    output_per_1k: 0.015
+    cache_read_per_1k: 0.0003
+  kimi-k2-thinking:
+    input_per_1k: 0.0006
+    output_per_1k: 0.0025
+    cache_read_per_1k: 0.00015
+
 bedrock:
   # AWS Bedrock model ids, normalised by the request parser (cross-region
   # inference-profile prefix + version/throughput suffix stripped), e.g.

proxy/internal/middleware/builtin/llm_guardrail/middleware.go

@@ -25,6 +25,14 @@ const (
 	denyCodeModel    = "llm_policy.model_blocked"
 	denyReasonModel  = "model_blocked"
 	denyMessageModel = "model is not in the policy allowlist"
+	// Deny reason used when an allowlist is configured but the request model
+	// could not be determined. URL/path-routed providers (AWS Bedrock, Google
+	// Vertex, ...) carry the model outside the JSON body, so a request shape the
+	// parser does not recognise reaches the guardrail with no model. Such a
+	// request must be denied (fail closed), never waved through.
+	denyCodeModelUnknown    = "llm_policy.model_unknown"
+	denyReasonModelUnknown  = "model_unknown"
+	denyMessageModelUnknown = "request model could not be determined for the policy allowlist"
 )
 
 // Middleware enforces the model allowlist and optionally captures the
@@ -108,23 +116,37 @@ func (m *Middleware) evaluateAllowlist(model string, modelPresent bool) *middlew
 	if len(m.cfg.ModelAllowlist) == 0 {
 		return nil
 	}
-	if !modelPresent {
-		return nil
+	// Fail closed: with an allowlist configured, a request whose model the
+	// upstream parser could not extract (absent or empty) must be denied rather
+	// than allowed. This is what enforces the allowlist for URL/path-routed
+	// providers (Bedrock, Vertex, ...) whose model lives outside the JSON body.
+	if !modelPresent || normaliseModel(model) == "" {
+		return denyModel("", denyCodeModelUnknown, denyMessageModelUnknown, denyReasonModelUnknown)
 	}
 	if m.modelInAllowlist(model) {
 		return nil
 	}
+	return denyModel(model, denyCodeModel, denyMessageModel, denyReasonModel)
+}
+
+// denyModel builds a 403 deny Output for a model-allowlist rejection. model is
+// included in the details only when non-empty.
+func denyModel(model, code, message, reason string) *middleware.Output {
+	details := map[string]string{}
+	if model != "" {
+		details["model"] = model
+	}
 	return &middleware.Output{
 		Decision:   middleware.DecisionDeny,
 		DenyStatus: 403,
 		DenyReason: &middleware.DenyReason{
-			Code:    denyCodeModel,
-			Message: denyMessageModel,
-			Details: map[string]string{"model": model},
+			Code:    code,
+			Message: message,
+			Details: details,
 		},
 		Metadata: []middleware.KV{
 			{Key: middleware.KeyLLMPolicyDecision, Value: "deny"},
-			{Key: middleware.KeyLLMPolicyReason, Value: denyReasonModel},
+			{Key: middleware.KeyLLMPolicyReason, Value: reason},
 		},
 	}
 }

proxy/internal/middleware/builtin/llm_guardrail/middleware_test.go

@@ -102,13 +102,44 @@ func TestAllowlistCaseInsensitive(t *testing.T) {
 	}
 }
 
-func TestAllowlistMissingModelKeyAllows(t *testing.T) {
+func TestAllowlistMissingModelKeyDenies(t *testing.T) {
+	// Fail closed: with an allowlist configured, a request whose model the
+	// parser could not extract (URL/path-routed providers such as Bedrock or
+	// Vertex whose shape wasn't recognised) must be denied, not allowed.
 	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
 	out, err := mw.Invoke(context.Background(), newInput())
 	require.NoError(t, err)
-	assert.Equal(t, middleware.DecisionAllow, out.Decision, "missing model key must allow even with non-empty allowlist")
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "absent model must be denied when an allowlist is set")
+	assert.Equal(t, 403, out.DenyStatus, "deny status must be 403")
+	require.NotNil(t, out.DenyReason, "deny reason must be populated")
+	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
 	dec, _ := metaValue(t, out.Metadata, middleware.KeyLLMPolicyDecision)
-	assert.Equal(t, "allow", dec, "decision must be allow when model key is absent")
+	assert.Equal(t, "deny", dec, "decision must be deny when model key is absent")
+	reason, _ := metaValue(t, out.Metadata, middleware.KeyLLMPolicyReason)
+	assert.Equal(t, "model_unknown", reason, "reason metadata must be model_unknown")
+}
+
+func TestAllowlistEmptyModelValueDenies(t *testing.T) {
+	// A present-but-empty model is as undeterminable as an absent one.
+	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
+	out, err := mw.Invoke(context.Background(), newInput(
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "   "},
+	))
+	require.NoError(t, err)
+	require.NotNil(t, out)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision, "empty model must be denied when an allowlist is set")
+	require.NotNil(t, out.DenyReason, "deny reason must be populated")
+	assert.Equal(t, "llm_policy.model_unknown", out.DenyReason.Code, "deny code must be model_unknown")
+}
+
+func TestAllowlistEmptyListAllowsMissingModel(t *testing.T) {
+	// Without an allowlist there is nothing to enforce, so a missing model is
+	// still allowed — the fail-closed rule only applies when a list is set.
+	mw := New(Config{})
+	out, err := mw.Invoke(context.Background(), newInput())
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision, "no allowlist must allow even without a model")
 }
 
 func TestPromptCaptureDisabledEmitsNoPrompt(t *testing.T) {

proxy/internal/middleware/builtin/llm_identity_inject/factory.go

@@ -64,6 +64,11 @@ type JSONMetadataRule struct {
 	UserKey        string `json:"user_key,omitempty"`
 	GroupsKey      string `json:"groups_key,omitempty"`
 	MaxValueLength int    `json:"max_value_length,omitempty"`
+	// Sanitize replaces characters outside the destination provider's accepted
+	// set with '_' before emitting each value. AWS Bedrock's
+	// X-Amzn-Bedrock-Request-Metadata restricts values to [A-Za-z0-9 +-=._:/@];
+	// group display names with other characters would otherwise 400.
+	Sanitize bool `json:"sanitize,omitempty"`
 }
 
 // Config is the on-wire configuration accepted by the factory. An

proxy/internal/middleware/builtin/llm_identity_inject/middleware.go

@@ -292,15 +292,21 @@ func applyJSONMetadata(rule *JSONMetadataRule, in *middleware.Input) *middleware
 	mutations := &middleware.Mutations{}
 	mutations.HeadersRemove = append(mutations.HeadersRemove, rule.Header)
 
+	emit := func(v string) string {
+		if rule.Sanitize {
+			v = sanitizeMetadataValue(v)
+		}
+		return truncate(v, rule.MaxValueLength)
+	}
 	payload := map[string]string{}
 	if rule.UserKey != "" {
 		if identity := identityFor(in); identity != "" {
-			payload[rule.UserKey] = truncate(identity, rule.MaxValueLength)
+			payload[rule.UserKey] = emit(identity)
 		}
 	}
 	if rule.GroupsKey != "" {
 		if csv := authorisingTagsCSV(in); csv != "" {
-			payload[rule.GroupsKey] = truncate(csv, rule.MaxValueLength)
+			payload[rule.GroupsKey] = emit(csv)
 		}
 	}
 	if len(payload) == 0 {
@@ -359,6 +365,36 @@ func truncate(s string, maxBytes int) string {
 	return s[:maxBytes]
 }
 
+// sanitizeMetadataValue replaces any character outside AWS Bedrock's accepted
+// request-metadata class — letters, digits, space, and + - = . _ : / @ — with
+// '_'. This keeps values (notably the groups CSV, whose commas are rejected, and
+// group display names with arbitrary characters) from making Bedrock reject the
+// request with 400. The result stays opaque to the gateway.
+func sanitizeMetadataValue(s string) string {
+	var b strings.Builder
+	b.Grow(len(s))
+	for _, r := range s {
+		if metadataCharAllowed(r) {
+			b.WriteRune(r)
+		} else {
+			b.WriteByte('_')
+		}
+	}
+	return b.String()
+}
+
+func metadataCharAllowed(r rune) bool {
+	switch {
+	case r >= 'a' && r <= 'z', r >= 'A' && r <= 'Z', r >= '0' && r <= '9':
+		return true
+	}
+	switch r {
+	case ' ', '+', '-', '=', '.', '_', ':', '/', '@':
+		return true
+	}
+	return false
+}
+
 // tagsIDsFromAuthorising reads llm_router's authorising-groups metadata
 // (a CSV of group ids) and returns the parsed slice. Returns nil when
 // the key is absent or empty so the caller can fall back to the full

proxy/internal/middleware/builtin/llm_identity_inject/middleware_test.go

@@ -304,6 +304,46 @@ func TestInject_JSONMetadata_TruncatesValues(t *testing.T) {
 		"per-value byte length must be capped at MaxValueLength")
 }
 
+// TestInject_JSONMetadata_Sanitize pins the AWS-Bedrock sanitization path: when
+// Sanitize is set, characters outside Bedrock's accepted metadata class
+// (notably the groups CSV comma and arbitrary characters in group display
+// names) are replaced with '_' so Bedrock doesn't reject the request. Allowed
+// characters (letters, digits, spaces, and @ . _ : / + - =) pass through.
+func TestInject_JSONMetadata_Sanitize(t *testing.T) {
+	rule := ProviderInjection{
+		ProviderID: portkeyProvider,
+		JSONMetadata: &JSONMetadataRule{
+			Header:         "X-Amzn-Bedrock-Request-Metadata",
+			UserKey:        "user",
+			GroupsKey:      "group",
+			MaxValueLength: 256,
+			Sanitize:       true,
+		},
+	}
+	mw := New(Config{Providers: []ProviderInjection{rule}})
+	in := newInput(portkeyProvider, "alice", []string{"g1", "g2"})
+	in.UserEmail = "alice@example.com"
+	// Group display names carry characters Bedrock rejects (comma, '#'); the CSV
+	// join adds another comma between the two groups.
+	in.UserGroupNames = []string{"Eng,Team", "Ops#1"}
+
+	out, err := mw.Invoke(context.Background(), in)
+	require.NoError(t, err)
+	require.NotNil(t, out.Mutations)
+	require.Len(t, out.Mutations.HeadersAdd, 1)
+	added := out.Mutations.HeadersAdd[0]
+	assert.Equal(t, "X-Amzn-Bedrock-Request-Metadata", added.Key,
+		"the Bedrock cost-allocation header carries the metadata JSON")
+
+	var payload map[string]string
+	require.NoError(t, json.Unmarshal([]byte(added.Value), &payload))
+	assert.Equal(t, "alice@example.com", payload["user"],
+		"'@' and '.' are in Bedrock's accepted set and must be preserved")
+	assert.NotContains(t, payload["group"], ",", "commas must be sanitized — Bedrock rejects them")
+	assert.NotContains(t, payload["group"], "#", "disallowed characters must be sanitized")
+	assert.Contains(t, payload["group"], "Eng", "allowed characters must be preserved")
+}
+
 // TestInject_JSONMetadata_EmptyIdentity_StripsButDoesNotAdd verifies the
 // anti-spoof Remove still fires when there's nothing to stamp.
 func TestInject_JSONMetadata_EmptyIdentity_StripsButDoesNotAdd(t *testing.T) {

proxy/internal/middleware/builtin/llm_request_parser/guardrail_allowlist_test.go

@@ -0,0 +1,106 @@
+package llm_request_parser
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_guardrail"
+)
+
+// runParserGuardrail runs the request parser then the model-allowlist guardrail
+// in SlotOnRequest order, threading the parser's metadata into the guardrail the
+// same way the real chain does. It returns the guardrail decision so tests can
+// assert allowlist enforcement for URL/path-routed providers end to end.
+func runParserGuardrail(t *testing.T, url string, body []byte, allowlist []string) *middleware.Output {
+	t.Helper()
+	parser := newMiddleware(t)
+	parsed, err := parser.Invoke(context.Background(), &middleware.Input{
+		Slot: middleware.SlotOnRequest,
+		URL:  url,
+		Body: body,
+	})
+	require.NoError(t, err, "parser must not error")
+
+	guard := llm_guardrail.New(llm_guardrail.Config{ModelAllowlist: allowlist})
+	out, err := guard.Invoke(context.Background(), &middleware.Input{
+		Slot:     middleware.SlotOnRequest,
+		Metadata: parsed.Metadata,
+	})
+	require.NoError(t, err, "guardrail must not error")
+	require.NotNil(t, out, "guardrail must return an output")
+	return out
+}
+
+// TestModelAllowlist_URLRoutedProviders validates that the model allowlist is
+// enforced for providers whose model travels in the URL path (AWS Bedrock,
+// Google Vertex) rather than the JSON body. The "unknown action" case is the
+// regression guard for #6751: a Bedrock request shape the parser cannot map to a
+// model must fail closed under an allowlist instead of bypassing it.
+func TestModelAllowlist_URLRoutedProviders(t *testing.T) {
+	const bedrockBody = `{"anthropic_version":"bedrock-2023-05-31","messages":[{"role":"user","content":"hi"}]}`
+	const vertexBody = `{"anthropic_version":"vertex-2023-10-16","messages":[{"role":"user","content":"hi"}]}`
+
+	tests := []struct {
+		name      string
+		url       string
+		body      string
+		allowlist []string
+		decision  middleware.Decision
+		denyCode  string
+	}{
+		{
+			name:      "bedrock allowed model passes",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-haiku-4-5-v1:0/invoke",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionAllow,
+		},
+		{
+			name:      "bedrock disallowed model denied",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-opus-4-8-v1:0/invoke",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+		{
+			name:      "bedrock unknown action fails closed",
+			url:       "https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-opus-4-8-v1:0/some-future-action",
+			body:      bedrockBody,
+			allowlist: []string{"anthropic.claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_unknown",
+		},
+		{
+			name:      "vertex disallowed model denied",
+			url:       "/v1/projects/p/locations/global/publishers/anthropic/models/claude-opus-4-8@20250101:rawPredict",
+			body:      vertexBody,
+			allowlist: []string{"claude-haiku-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+		{
+			name:      "vertex allowed model passes",
+			url:       "/v1/projects/p/locations/global/publishers/anthropic/models/claude-haiku-4-5@20250101:rawPredict",
+			body:      vertexBody,
+			allowlist: []string{"claude-haiku-4-5"},
+			decision:  middleware.DecisionAllow,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			out := runParserGuardrail(t, tt.url, []byte(tt.body), tt.allowlist)
+			assert.Equal(t, tt.decision, out.Decision, "unexpected decision for %s", tt.name)
+			if tt.decision == middleware.DecisionDeny {
+				require.NotNil(t, out.DenyReason, "deny reason must be set for %s", tt.name)
+				assert.Equal(t, 403, out.DenyStatus, "deny status must be 403 for %s", tt.name)
+				assert.Equal(t, tt.denyCode, out.DenyReason.Code, "deny code for %s", tt.name)
+			}
+		})
+	}
+}

proxy/internal/middleware/builtin/llm_router/bedrock_route_test.go

@@ -0,0 +1,30 @@
+package llm_router
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+)
+
+// TestRouteClaimsModel_BedrockNormalizesCandidate guards the fix for the native
+// Bedrock routing gap: the request model reaches the router already normalized
+// (the parser strips the region/inference-profile prefix and version suffix),
+// so a provider registered with the raw inference-profile id must still match.
+func TestRouteClaimsModel_BedrockNormalizesCandidate(t *testing.T) {
+	route := ProviderRoute{Bedrock: true, Models: []string{"us.anthropic.claude-haiku-4-5"}}
+	assert.True(t, routeClaimsModel(route, "anthropic.claude-haiku-4-5"),
+		"raw region-prefixed Bedrock model must match the normalized request model")
+	assert.False(t, routeClaimsModel(route, "anthropic.claude-opus-4-8"),
+		"a model outside the provider's list must not match")
+
+	// A provider registered with the already-normalized id also matches.
+	normalized := ProviderRoute{Bedrock: true, Models: []string{"anthropic.claude-haiku-4-5"}}
+	assert.True(t, routeClaimsModel(normalized, "anthropic.claude-haiku-4-5"),
+		"normalized Bedrock model must match")
+
+	// Non-Bedrock routes keep exact matching (no prefix stripping).
+	openai := ProviderRoute{Models: []string{"gpt-4o"}}
+	assert.True(t, routeClaimsModel(openai, "gpt-4o"), "exact model must match")
+	assert.False(t, routeClaimsModel(openai, "us.gpt-4o"),
+		"non-Bedrock routes must not strip a us. prefix")
+}

proxy/internal/middleware/builtin/llm_router/middleware.go

@@ -23,6 +23,7 @@ import (
 	"golang.org/x/oauth2"
 	"golang.org/x/oauth2/google"
 
+	"github.com/netbirdio/netbird/proxy/internal/llm"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
 )
 
@@ -555,6 +556,14 @@ func routeClaimsModel(route ProviderRoute, model string) bool {
 		if candidate == model {
 			return true
 		}
+		// Bedrock request models reach the router already normalized (the parser
+		// strips the region / inference-profile prefix and version suffix), but
+		// the operator may register the raw inference-profile id (e.g.
+		// "us.anthropic.claude-haiku-4-5"). Normalize the candidate so both sides
+		// compare equal; otherwise a native Bedrock request denies as not-routable.
+		if route.Bedrock && llm.NormalizeBedrockModel(candidate) == model {
+			return true
+		}
 	}
 	return false
 }

proxy/internal/proxy/reverseproxy.go

@@ -22,6 +22,7 @@ import (
 	"github.com/netbirdio/netbird/proxy/internal/roundtrip"
 	"github.com/netbirdio/netbird/proxy/internal/types"
 	"github.com/netbirdio/netbird/proxy/web"
+	"github.com/netbirdio/netbird/trustedproxy"
 )
 
 type ReverseProxy struct {
@@ -29,10 +30,10 @@ type ReverseProxy struct {
 	// forwardedProto overrides the X-Forwarded-Proto header value.
 	// Valid values: "auto" (detect from TLS), "http", "https".
 	forwardedProto string
-	// trustedProxies is a list of IP prefixes for trusted upstream proxies.
-	// When the direct connection comes from a trusted proxy, forwarding
-	// headers are preserved and appended to instead of being stripped.
-	trustedProxies []netip.Prefix
+	// trustedProxies is the set of trusted upstream proxies. When the direct
+	// connection comes from a trusted proxy, forwarding headers are preserved
+	// and appended to instead of being stripped.
+	trustedProxies *trustedproxy.List
 	mappingsMux    sync.RWMutex
 	mappings       map[string]Mapping
 	logger         *log.Logger
@@ -63,7 +64,7 @@ func WithMiddlewareManager(m *middleware.Manager) Option {
 // between requested URLs and targets.
 // The internal mappings can be modified using the AddMapping
 // and RemoveMapping functions.
-func NewReverseProxy(transport http.RoundTripper, forwardedProto string, trustedProxies []netip.Prefix, logger *log.Logger, opts ...Option) *ReverseProxy {
+func NewReverseProxy(transport http.RoundTripper, forwardedProto string, trustedProxies *trustedproxy.List, logger *log.Logger, opts ...Option) *ReverseProxy {
 	if logger == nil {
 		logger = log.StandardLogger()
 	}
@@ -527,7 +528,7 @@ func (p *ReverseProxy) isSelfTargetLoop(r *http.Request, target *url.URL) bool {
 	if !types.IsOverlayOrigin(r.Context()) {
 		return false
 	}
-	srcIP := extractHostIP(r.RemoteAddr)
+	srcIP := trustedproxy.ExtractHostIP(r.RemoteAddr)
 	if !srcIP.IsValid() {
 		return false
 	}
@@ -578,9 +579,9 @@ func (p *ReverseProxy) rewriteFunc(target *url.URL, matchedPath string, passHost
 
 		stampNetBirdIdentity(r)
 
-		clientIP := extractHostIP(r.In.RemoteAddr)
+		clientIP := trustedproxy.ExtractHostIP(r.In.RemoteAddr)
 
-		if isTrustedAddr(clientIP, p.trustedProxies) {
+		if p.trustedProxies.Contains(clientIP) {
 			p.setTrustedForwardingHeaders(r, clientIP)
 		} else {
 			p.setUntrustedForwardingHeaders(r, clientIP)
@@ -664,7 +665,7 @@ func (p *ReverseProxy) setTrustedForwardingHeaders(r *httputil.ProxyRequest, cli
 	if realIP := r.In.Header.Get("X-Real-IP"); realIP != "" {
 		r.Out.Header.Set("X-Real-IP", realIP)
 	} else {
-		resolved := ResolveClientIP(r.In.RemoteAddr, r.In.Header.Get("X-Forwarded-For"), p.trustedProxies)
+		resolved := p.trustedProxies.ResolveClientIP(r.In.RemoteAddr, r.In.Header.Get("X-Forwarded-For"))
 		r.Out.Header.Set("X-Real-IP", resolved.String())
 	}
 

proxy/internal/proxy/reverseproxy_test.go

@@ -23,6 +23,7 @@ import (
 	"github.com/netbirdio/netbird/proxy/internal/roundtrip"
 	"github.com/netbirdio/netbird/proxy/internal/types"
 	"github.com/netbirdio/netbird/proxy/web"
+	"github.com/netbirdio/netbird/trustedproxy"
 )
 
 func TestRewriteFunc_HostRewriting(t *testing.T) {
@@ -302,7 +303,7 @@ func TestExtractHostIP(t *testing.T) {
 	}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
-			assert.Equal(t, tt.expected, extractHostIP(tt.remoteAddr))
+			assert.Equal(t, tt.expected, trustedproxy.ExtractHostIP(tt.remoteAddr))
 		})
 	}
 }
@@ -330,7 +331,7 @@ func TestExtractForwardedPort(t *testing.T) {
 
 func TestRewriteFunc_TrustedProxy(t *testing.T) {
 	target, _ := url.Parse("http://backend.internal:8080")
-	trusted := []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")}
+	trusted := trustedproxy.FromPrefixes([]netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")})
 
 	t.Run("appends to X-Forwarded-For", func(t *testing.T) {
 		p := &ReverseProxy{forwardedProto: "auto", trustedProxies: trusted}

proxy/internal/proxy/trustedproxy.go

@@ -1,81 +0,0 @@
-package proxy
-
-import (
-	"net/netip"
-	"strings"
-)
-
-// IsTrustedProxy checks if the given IP string falls within any of the trusted prefixes.
-func IsTrustedProxy(ipStr string, trusted []netip.Prefix) bool {
-	addr, err := netip.ParseAddr(ipStr)
-	if err != nil || len(trusted) == 0 {
-		return false
-	}
-	return isTrustedAddr(addr.Unmap(), trusted)
-}
-
-// ResolveClientIP extracts the real client IP from X-Forwarded-For using the trusted proxy list.
-// It walks the XFF chain right-to-left, skipping IPs that match trusted prefixes.
-// The first untrusted IP is the real client.
-//
-// If the trusted list is empty or remoteAddr is not trusted, it returns the
-// remoteAddr IP directly (ignoring any forwarding headers).
-func ResolveClientIP(remoteAddr, xff string, trusted []netip.Prefix) netip.Addr {
-	remoteIP := extractHostIP(remoteAddr)
-
-	if len(trusted) == 0 || !isTrustedAddr(remoteIP, trusted) {
-		return remoteIP
-	}
-
-	if xff == "" {
-		return remoteIP
-	}
-
-	parts := strings.Split(xff, ",")
-	for i := len(parts) - 1; i >= 0; i-- {
-		ip := strings.TrimSpace(parts[i])
-		if ip == "" {
-			continue
-		}
-		addr, err := netip.ParseAddr(ip)
-		if err != nil {
-			continue
-		}
-		addr = addr.Unmap()
-		if !isTrustedAddr(addr, trusted) {
-			return addr
-		}
-	}
-
-	// All IPs in XFF are trusted; return the leftmost as best guess.
-	if first := strings.TrimSpace(parts[0]); first != "" {
-		if addr, err := netip.ParseAddr(first); err == nil {
-			return addr.Unmap()
-		}
-	}
-	return remoteIP
-}
-
-// extractHostIP parses the IP from a host:port string and returns it unmapped.
-func extractHostIP(hostPort string) netip.Addr {
-	if ap, err := netip.ParseAddrPort(hostPort); err == nil {
-		return ap.Addr().Unmap()
-	}
-	if addr, err := netip.ParseAddr(hostPort); err == nil {
-		return addr.Unmap()
-	}
-	return netip.Addr{}
-}
-
-// isTrustedAddr checks if the given address falls within any of the trusted prefixes.
-func isTrustedAddr(addr netip.Addr, trusted []netip.Prefix) bool {
-	if !addr.IsValid() {
-		return false
-	}
-	for _, prefix := range trusted {
-		if prefix.Contains(addr) {
-			return true
-		}
-	}
-	return false
-}

proxy/internal/proxy/trustedproxy_test.go

@@ -1,129 +0,0 @@
-package proxy
-
-import (
-	"net/netip"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-)
-
-func TestIsTrustedProxy(t *testing.T) {
-	trusted := []netip.Prefix{
-		netip.MustParsePrefix("10.0.0.0/8"),
-		netip.MustParsePrefix("192.168.1.0/24"),
-		netip.MustParsePrefix("fd00::/8"),
-	}
-
-	tests := []struct {
-		name    string
-		ip      string
-		trusted []netip.Prefix
-		want    bool
-	}{
-		{"empty trusted list", "10.0.0.1", nil, false},
-		{"IP within /8 prefix", "10.1.2.3", trusted, true},
-		{"IP within /24 prefix", "192.168.1.100", trusted, true},
-		{"IP outside all prefixes", "203.0.113.50", trusted, false},
-		{"boundary IP just outside prefix", "192.168.2.1", trusted, false},
-		{"unparsable IP", "not-an-ip", trusted, false},
-		{"IPv6 in trusted range", "fd00::1", trusted, true},
-		{"IPv6 outside range", "2001:db8::1", trusted, false},
-		{"empty string", "", trusted, false},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			assert.Equal(t, tt.want, IsTrustedProxy(tt.ip, tt.trusted))
-		})
-	}
-}
-
-func TestResolveClientIP(t *testing.T) {
-	trusted := []netip.Prefix{
-		netip.MustParsePrefix("10.0.0.0/8"),
-		netip.MustParsePrefix("172.16.0.0/12"),
-	}
-
-	tests := []struct {
-		name       string
-		remoteAddr string
-		xff        string
-		trusted    []netip.Prefix
-		want       netip.Addr
-	}{
-		{
-			name:       "empty trusted list returns RemoteAddr",
-			remoteAddr: "203.0.113.50:9999",
-			xff:        "1.2.3.4",
-			trusted:    nil,
-			want:       netip.MustParseAddr("203.0.113.50"),
-		},
-		{
-			name:       "untrusted RemoteAddr ignores XFF",
-			remoteAddr: "203.0.113.50:9999",
-			xff:        "1.2.3.4, 10.0.0.1",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("203.0.113.50"),
-		},
-		{
-			name:       "trusted RemoteAddr with single client in XFF",
-			remoteAddr: "10.0.0.1:5000",
-			xff:        "203.0.113.50",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("203.0.113.50"),
-		},
-		{
-			name:       "trusted RemoteAddr walks past trusted entries in XFF",
-			remoteAddr: "10.0.0.1:5000",
-			xff:        "203.0.113.50, 10.0.0.2, 172.16.0.5",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("203.0.113.50"),
-		},
-		{
-			name:       "trusted RemoteAddr with empty XFF falls back to RemoteAddr",
-			remoteAddr: "10.0.0.1:5000",
-			xff:        "",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("10.0.0.1"),
-		},
-		{
-			name:       "all XFF IPs trusted returns leftmost",
-			remoteAddr: "10.0.0.1:5000",
-			xff:        "10.0.0.2, 172.16.0.1, 10.0.0.3",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("10.0.0.2"),
-		},
-		{
-			name:       "XFF with whitespace",
-			remoteAddr: "10.0.0.1:5000",
-			xff:        " 203.0.113.50 , 10.0.0.2 ",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("203.0.113.50"),
-		},
-		{
-			name:       "XFF with empty segments",
-			remoteAddr: "10.0.0.1:5000",
-			xff:        "203.0.113.50,,10.0.0.2",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("203.0.113.50"),
-		},
-		{
-			name:       "multi-hop with mixed trust",
-			remoteAddr: "10.0.0.1:5000",
-			xff:        "8.8.8.8, 203.0.113.50, 172.16.0.1",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("203.0.113.50"),
-		},
-		{
-			name:       "RemoteAddr without port",
-			remoteAddr: "10.0.0.1",
-			xff:        "203.0.113.50",
-			trusted:    trusted,
-			want:       netip.MustParseAddr("203.0.113.50"),
-		},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			assert.Equal(t, tt.want, ResolveClientIP(tt.remoteAddr, tt.xff, tt.trusted))
-		})
-	}
-}

proxy/lifecycle.go

@@ -2,13 +2,13 @@ package proxy
 
 import (
 	"context"
-	"net/netip"
 	"time"
 
 	log "github.com/sirupsen/logrus"
 
 	"github.com/netbirdio/netbird/client/embed"
 	"github.com/netbirdio/netbird/proxy/internal/acme"
+	"github.com/netbirdio/netbird/trustedproxy"
 )
 
 // Config bundles every knob the proxy reads at construction time. It mirrors
@@ -83,9 +83,9 @@ type Config struct {
 	// ForwardedProto overrides the X-Forwarded-Proto value sent to
 	// backends. Valid values: "auto", "http", "https".
 	ForwardedProto string
-	// TrustedProxies is a list of IP prefixes for trusted upstream
-	// proxies that may set forwarding headers.
-	TrustedProxies []netip.Prefix
+	// TrustedProxies is the set of trusted upstream proxies that may set
+	// forwarding headers.
+	TrustedProxies *trustedproxy.List
 	// WireguardPort is the UDP port for the embedded NetBird tunnel.
 	// Zero asks the OS for a random port.
 	WireguardPort uint16

proxy/proxyprotocol_test.go

@@ -10,12 +10,14 @@ import (
 	log "github.com/sirupsen/logrus"
 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/trustedproxy"
 )
 
 func TestWrapProxyProtocol_OverridesRemoteAddr(t *testing.T) {
 	srv := &Server{
 		Logger:         log.StandardLogger(),
-		TrustedProxies: []netip.Prefix{netip.MustParsePrefix("127.0.0.1/32")},
+		TrustedProxies: trustedproxy.FromPrefixes([]netip.Prefix{netip.MustParsePrefix("127.0.0.1/32")}),
 		ProxyProtocol:  true,
 	}
 
@@ -66,7 +68,7 @@ func TestWrapProxyProtocol_OverridesRemoteAddr(t *testing.T) {
 func TestProxyProtocolPolicy_TrustedRequires(t *testing.T) {
 	srv := &Server{
 		Logger:         log.StandardLogger(),
-		TrustedProxies: []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")},
+		TrustedProxies: trustedproxy.FromPrefixes([]netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")}),
 	}
 
 	opts := proxyproto.ConnPolicyOptions{
@@ -80,7 +82,7 @@ func TestProxyProtocolPolicy_TrustedRequires(t *testing.T) {
 func TestProxyProtocolPolicy_UntrustedIgnores(t *testing.T) {
 	srv := &Server{
 		Logger:         log.StandardLogger(),
-		TrustedProxies: []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")},
+		TrustedProxies: trustedproxy.FromPrefixes([]netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")}),
 	}
 
 	opts := proxyproto.ConnPolicyOptions{
@@ -94,7 +96,7 @@ func TestProxyProtocolPolicy_UntrustedIgnores(t *testing.T) {
 func TestProxyProtocolPolicy_InvalidIPRejects(t *testing.T) {
 	srv := &Server{
 		Logger:         log.StandardLogger(),
-		TrustedProxies: []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")},
+		TrustedProxies: trustedproxy.FromPrefixes([]netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")}),
 	}
 
 	opts := proxyproto.ConnPolicyOptions{

proxy/server.go

@@ -67,6 +67,7 @@ import (
 	"github.com/netbirdio/netbird/proxy/web"
 	"github.com/netbirdio/netbird/shared/management/domain"
 	"github.com/netbirdio/netbird/shared/management/proto"
+	"github.com/netbirdio/netbird/trustedproxy"
 	"github.com/netbirdio/netbird/util/embeddedroots"
 )
 
@@ -79,19 +80,19 @@ type portRouter struct {
 
 type Server struct {
 	ctx               context.Context
-	mgmtClient    proto.ProxyServiceClient
-	proxy         *proxy.ReverseProxy
-	netbird       *roundtrip.NetBird
-	acme          *acme.Manager
+	mgmtClient        proto.ProxyServiceClient
+	proxy             *proxy.ReverseProxy
+	netbird           *roundtrip.NetBird
+	acme              *acme.Manager
 	staticCertWatcher *certwatch.Watcher
-	auth          *auth.Middleware
-	http          *http.Server
-	https         *http.Server
-	debug         *http.Server
-	healthServer  *health.Server
-	healthChecker *health.Checker
-	meter         *proxymetrics.Metrics
-	accessLog     *accesslog.Logger
+	auth              *auth.Middleware
+	http              *http.Server
+	https             *http.Server
+	debug             *http.Server
+	healthServer      *health.Server
+	healthChecker     *health.Checker
+	meter             *proxymetrics.Metrics
+	accessLog         *accesslog.Logger
 	// middlewareManager drives per-target middleware dispatch. Always
 	// constructed during boot; an empty registry produces empty chains and
 	// the reverse-proxy stays on the no-capture fast path.
@@ -99,16 +100,16 @@ type Server struct {
 	// middlewareRegistry is the source of registered middleware factories.
 	// Concrete middlewares register themselves through init().
 	middlewareRegistry *middleware.Registry
-	mainRouter    *nbtcp.Router
-	mainPort      uint16
-	udpMu         sync.Mutex
-	udpRelays     map[types.ServiceID]*udprelay.Relay
-	udpRelayWg    sync.WaitGroup
-	portMu        sync.RWMutex
-	portRouters   map[uint16]*portRouter
-	svcPorts      map[types.ServiceID][]uint16
-	lastMappings  map[types.ServiceID]*proto.ProxyMapping
-	portRouterWg  sync.WaitGroup
+	mainRouter         *nbtcp.Router
+	mainPort           uint16
+	udpMu              sync.Mutex
+	udpRelays          map[types.ServiceID]*udprelay.Relay
+	udpRelayWg         sync.WaitGroup
+	portMu             sync.RWMutex
+	portRouters        map[uint16]*portRouter
+	svcPorts           map[types.ServiceID][]uint16
+	lastMappings       map[types.ServiceID]*proto.ProxyMapping
+	portRouterWg       sync.WaitGroup
 
 	// hijackTracker tracks hijacked connections (e.g. WebSocket upgrades)
 	// so they can be closed during graceful shutdown, since http.Server.Shutdown
@@ -192,10 +193,10 @@ type Server struct {
 	// ForwardedProto overrides the X-Forwarded-Proto value sent to backends.
 	// Valid values: "auto" (detect from TLS), "http", "https".
 	ForwardedProto string
-	// TrustedProxies is a list of IP prefixes for trusted upstream proxies.
-	// When set, forwarding headers from these sources are preserved and
-	// appended to instead of being stripped.
-	TrustedProxies []netip.Prefix
+	// TrustedProxies is the set of trusted upstream proxies. When set,
+	// forwarding headers from these sources are preserved and appended to
+	// instead of being stripped.
+	TrustedProxies *trustedproxy.List
 	// WireguardPort is the port for the NetBird tunnel interface. Use 0
 	// for a random OS-assigned port. A fixed port only works with
 	// single-account deployments; multiple accounts will fail to bind
@@ -718,7 +719,7 @@ func (s *Server) wrapProxyProtocol(ln net.Listener) net.Listener {
 		Listener:          ln,
 		ReadHeaderTimeout: proxyProtoHeaderTimeout,
 	}
-	if len(s.TrustedProxies) > 0 {
+	if !s.TrustedProxies.Empty() {
 		ppListener.ConnPolicy = s.proxyProtocolPolicy
 	} else {
 		s.Logger.Warn("PROXY protocol enabled without trusted proxies; any source may send PROXY headers")
@@ -742,10 +743,8 @@ func (s *Server) proxyProtocolPolicy(opts proxyproto.ConnPolicyOptions) (proxypr
 	addr = addr.Unmap()
 
 	// called per accept
-	for _, prefix := range s.TrustedProxies {
-		if prefix.Contains(addr) {
-			return proxyproto.REQUIRE, nil
-		}
+	if s.TrustedProxies.Contains(addr) {
+		return proxyproto.REQUIRE, nil
 	}
 	return proxyproto.IGNORE, nil
 }

proxy/trustedproxy.go

@@ -1,43 +0,0 @@
-package proxy
-
-import (
-	"fmt"
-	"net/netip"
-	"strings"
-)
-
-// ParseTrustedProxies parses a comma-separated list of CIDR prefixes or bare IPs
-// into a slice of netip.Prefix values suitable for trusted proxy configuration.
-// Bare IPs are converted to single-host prefixes (/32 or /128).
-func ParseTrustedProxies(raw string) ([]netip.Prefix, error) {
-	if raw == "" {
-		return nil, nil
-	}
-
-	parts := strings.Split(raw, ",")
-	prefixes := make([]netip.Prefix, 0, len(parts))
-	for _, part := range parts {
-		part = strings.TrimSpace(part)
-		if part == "" {
-			continue
-		}
-
-		prefix, err := netip.ParsePrefix(part)
-		if err == nil {
-			prefixes = append(prefixes, prefix)
-			continue
-		}
-
-		addr, addrErr := netip.ParseAddr(part)
-		if addrErr != nil {
-			return nil, fmt.Errorf("parse trusted proxy %q: not a valid CIDR or IP: %w", part, addrErr)
-		}
-
-		bits := 32
-		if addr.Is6() {
-			bits = 128
-		}
-		prefixes = append(prefixes, netip.PrefixFrom(addr, bits))
-	}
-	return prefixes, nil
-}