[proxy] full-chain regression test: Vertex model allowlist is enforced

Add a realstack test that drives Anthropic-on-Vertex requests (model in
the URL path, no model in the body — the shape Claude Code sends) through
the full synthesized middleware chain against an in-process management
stack and a fake Vertex upstream, reproducing the customer scenario end
to end.

With a guardrail allowlisting only Sonnet, an Opus rawPredict must be
denied by the guardrail (llm_policy.model_blocked) before the upstream,
and Sonnet must be served. Covers two provider shapes:

  - catch_all: provider registered with no models (historical Vertex
    default) — the router waves every model through, so this is the
    direct guard for "the disallowed model is never blocked".
  - versioned_models: provider registered with the raw "@version" ids —
    the router must normalize the registered id to route the
    version-stripped request model while the guardrail still blocks Opus.

A static bearer api key is used so the router injects a static auth
header instead of minting a GCP OAuth token (which needs network egress
and would deny before the guardrail runs, masking the decision under
test). The upstream-hit counter proves a denied request never leaks to
Google.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XDgMfwgU1Xerfaedu6bQse

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/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/conn_mgr.go

@@ -34,6 +34,8 @@ const (
 // - Handling connection establishment based on peer signaling
 //
 // The implementation is not thread-safe; it is protected by engine.syncMsgMux.
+// The only exception is ActivatePeer, which is safe for concurrent use so the
+// DNS warm-up path can call it without contending on the engine mutex.
 type ConnMgr struct {
 	peerStore        *peerstore.Store
 	statusRecorder   *peer.Status
@@ -42,6 +44,10 @@ type ConnMgr struct {
 	rosenpassEnabled bool
 
 	lazyConnMgr *manager.Manager
+	// lazyConnMgrMu guards the lazyConnMgr pointer for readers outside the
+	// engine loop (ActivatePeer). Writers hold it in addition to
+	// engine.syncMsgMux; all other reads stay under engine.syncMsgMux only.
+	lazyConnMgrMu sync.RWMutex
 
 	wg            sync.WaitGroup
 	lazyCtx       context.Context
@@ -238,12 +244,20 @@ func (e *ConnMgr) RemovePeerConn(peerKey string) {
 	conn.Log.Infof("removed peer from lazy conn manager")
 }
 
+// ActivatePeer wakes an idle lazy connection. Unlike the rest of ConnMgr it is
+// safe for concurrent use: the lazy manager pointer is read under lazyConnMgrMu
+// and the manager itself is internally synchronized, so callers outside the
+// engine loop (DNS warm-up) do not need engine.syncMsgMux.
 func (e *ConnMgr) ActivatePeer(ctx context.Context, conn *peer.Conn) {
-	if !e.isStartedWithLazyMgr() {
+	e.lazyConnMgrMu.RLock()
+	lazyConnMgr := e.lazyConnMgr
+	started := lazyConnMgr != nil && e.lazyCtxCancel != nil
+	e.lazyConnMgrMu.RUnlock()
+	if !started {
 		return
 	}
 
-	if found := e.lazyConnMgr.ActivatePeer(conn.GetKey()); found {
+	if found := lazyConnMgr.ActivatePeer(conn.GetKey()); found {
 		if err := conn.Open(ctx); err != nil {
 			conn.Log.Errorf("failed to open connection: %v", err)
 		}
@@ -268,16 +282,21 @@ func (e *ConnMgr) Close() {
 
 	e.lazyCtxCancel()
 	e.wg.Wait()
+
+	e.lazyConnMgrMu.Lock()
 	e.lazyConnMgr = nil
+	e.lazyConnMgrMu.Unlock()
 }
 
 func (e *ConnMgr) initLazyManager(engineCtx context.Context) {
 	cfg := manager.Config{
 		InactivityThreshold: inactivityThresholdEnv(),
 	}
-	e.lazyConnMgr = manager.NewManager(cfg, engineCtx, e.peerStore, e.iface)
 
+	e.lazyConnMgrMu.Lock()
+	e.lazyConnMgr = manager.NewManager(cfg, engineCtx, e.peerStore, e.iface)
 	e.lazyCtx, e.lazyCtxCancel = context.WithCancel(engineCtx)
+	e.lazyConnMgrMu.Unlock()
 
 	e.wg.Add(1)
 	go func() {
@@ -316,7 +335,10 @@ func (e *ConnMgr) closeManager(ctx context.Context) {
 
 	e.lazyCtxCancel()
 	e.wg.Wait()
+
+	e.lazyConnMgrMu.Lock()
 	e.lazyConnMgr = nil
+	e.lazyConnMgrMu.Unlock()
 
 	for _, peerID := range e.peerStore.PeersPubKey() {
 		e.peerStore.PeerConnOpen(ctx, peerID)

client/internal/conn_mgr_test.go

@@ -1,10 +1,21 @@
 package internal
 
 import (
+	"context"
+	"net"
+	"net/netip"
 	"os"
+	"sync"
 	"testing"
+	"time"
 
+	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
+
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
 	"github.com/netbirdio/netbird/client/internal/lazyconn"
+	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/client/internal/peerstore"
+	"github.com/netbirdio/netbird/monotime"
 )
 
 func TestResolveLazyForce(t *testing.T) {
@@ -38,3 +49,58 @@ func TestResolveLazyForce(t *testing.T) {
 		})
 	}
 }
+
+type mockLazyWGIface struct{}
+
+func (mockLazyWGIface) RemovePeer(string) error { return nil }
+func (mockLazyWGIface) UpdatePeer(string, []netip.Prefix, time.Duration, *net.UDPAddr, *wgtypes.Key) error {
+	return nil
+}
+func (mockLazyWGIface) IsUserspaceBind() bool                    { return false }
+func (mockLazyWGIface) Address() wgaddr.Address                  { return wgaddr.Address{} }
+func (mockLazyWGIface) LastActivities() map[string]monotime.Time { return nil }
+func (mockLazyWGIface) MTU() uint16                              { return 1280 }
+
+// TestConnMgr_ActivatePeerConcurrentWithLifecycle exercises ActivatePeer from
+// non-engine goroutines (the DNS warm-up path) racing the manager lifecycle,
+// which stays on the engine loop. Run with -race: it fails if ActivatePeer
+// still requires engine.syncMsgMux for safety.
+func TestConnMgr_ActivatePeerConcurrentWithLifecycle(t *testing.T) {
+	t.Setenv(lazyconn.EnvLazyConn, "on")
+
+	status := peer.NewRecorder("https://mgm")
+	store := peerstore.NewConnStore()
+	connMgr := NewConnMgr(&EngineConfig{}, status, store, mockLazyWGIface{})
+
+	conn := newTestPeerConn(t, "peerA")
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	connMgr.Start(ctx)
+
+	done := make(chan struct{})
+	var wg sync.WaitGroup
+	for range 4 {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			for {
+				select {
+				case <-done:
+					return
+				default:
+					connMgr.ActivatePeer(ctx, conn)
+				}
+			}
+		}()
+	}
+
+	// Let the activators spin against the started manager, then tear it down
+	// underneath them and let them spin against the stopped manager.
+	time.Sleep(100 * time.Millisecond)
+	connMgr.Close()
+	time.Sleep(50 * time.Millisecond)
+
+	close(done)
+	wg.Wait()
+}

client/internal/dns/local/local.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"net"
 	"net/netip"
+	"os"
 	"slices"
 	"strings"
 	"sync"
@@ -36,7 +37,43 @@ type resolver interface {
 // record is left alone (it points at something outside our mesh, e.g.
 // a non-peer upstream).
 type PeerConnectivity interface {
-	IsConnectedByIP(ip string) (known, connected bool)
+	IsConnectedByIP(ip netip.Addr) (known, connected bool)
+}
+
+// PeerActivator wakes lazy-connection peers on demand. The local resolver calls
+// it with the tunnel IPs an answer points at, so a peer that is idle (lazily
+// disconnected) starts connecting at DNS-resolution time rather than racing the
+// client's first request packet. nil disables warm-up.
+type PeerActivator interface {
+	// ActivatePeersByIP triggers wake-up for the peer(s) owning addrs and blocks
+	// until one is connected or ctx (a short per-query budget) expires. It is a
+	// fast no-op for unknown or already-connected addresses.
+	ActivatePeersByIP(ctx context.Context, addrs []netip.Addr)
+}
+
+const (
+	defaultLazyWarmupTimeout = 2 * time.Second
+	envLazyWarmupTimeout     = "NB_DNS_LAZY_WARMUP_TIMEOUT"
+)
+
+// lazyWarmupTimeoutFromEnv returns the per-query budget for waking a
+// lazy-connection peer a DNS answer points at. Tunable via
+// NB_DNS_LAZY_WARMUP_TIMEOUT (a Go duration). Parsed once at construction time.
+func lazyWarmupTimeoutFromEnv() time.Duration {
+	v := os.Getenv(envLazyWarmupTimeout)
+	if v == "" {
+		return defaultLazyWarmupTimeout
+	}
+	d, err := time.ParseDuration(v)
+	if err != nil {
+		log.Warnf("invalid %s value %q, using default %s: %v", envLazyWarmupTimeout, v, defaultLazyWarmupTimeout, err)
+		return defaultLazyWarmupTimeout
+	}
+	if d <= 0 {
+		log.Warnf("non-positive %s value %q, using default %s", envLazyWarmupTimeout, v, defaultLazyWarmupTimeout)
+		return defaultLazyWarmupTimeout
+	}
+	return d
 }
 
 type Resolver struct {
@@ -51,6 +88,12 @@ type Resolver struct {
 	// filter and preserves the legacy "return whatever is registered"
 	// behaviour for callers that never wire a status source.
 	peerConn PeerConnectivity
+	// peerActivator, when non-nil, is called at resolution time to warm the
+	// lazy connection to the peer(s) an answer points at. nil disables warm-up.
+	peerActivator PeerActivator
+	// warmupTimeout is the per-query budget for the lazy-connection warm-up
+	// wait, resolved from the environment once at construction time.
+	warmupTimeout time.Duration
 
 	ctx    context.Context
 	cancel context.CancelFunc
@@ -59,11 +102,12 @@ type Resolver struct {
 func NewResolver() *Resolver {
 	ctx, cancel := context.WithCancel(context.Background())
 	return &Resolver{
-		records: make(map[dns.Question][]dns.RR),
-		domains: make(map[domain.Domain]struct{}),
-		zones:   make(map[domain.Domain]bool),
-		ctx:     ctx,
-		cancel:  cancel,
+		records:       make(map[dns.Question][]dns.RR),
+		domains:       make(map[domain.Domain]struct{}),
+		zones:         make(map[domain.Domain]bool),
+		warmupTimeout: lazyWarmupTimeoutFromEnv(),
+		ctx:           ctx,
+		cancel:        cancel,
 	}
 }
 
@@ -76,6 +120,14 @@ func (d *Resolver) SetPeerConnectivity(p PeerConnectivity) {
 	d.peerConn = p
 }
 
+// SetPeerActivator wires the DNS-time lazy-connection warm-up. Pass nil to
+// disable. Safe to call multiple times; the latest value wins.
+func (d *Resolver) SetPeerActivator(a PeerActivator) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	d.peerActivator = a
+}
+
 func (d *Resolver) MatchSubdomains() bool {
 	return true
 }
@@ -122,6 +174,9 @@ func (d *Resolver) ServeDNS(w dns.ResponseWriter, r *dns.Msg) {
 	replyMessage.RecursionAvailable = true
 
 	result := d.lookupRecords(logger, question)
+	// Warm before filtering: activation flips a lazily-idle target to connected,
+	// which then lets it survive the disconnected-peer filter below.
+	d.warmLazyPeers(question, result.records)
 	result.records = d.filterDisconnectedPeerAnswers(logger, question, result.records)
 	replyMessage.Authoritative = !result.hasExternalData
 	replyMessage.Answer = result.records
@@ -495,8 +550,8 @@ func (d *Resolver) filterDisconnectedPeerAnswers(logger *log.Entry, question dns
 	kept := make([]dns.RR, 0, len(records))
 	var dropped int
 	for _, rr := range records {
-		ip := extractRecordIP(rr)
-		if ip == "" {
+		ip, ok := extractRecordAddr(rr)
+		if !ok {
 			kept = append(kept, rr)
 			continue
 		}
@@ -518,22 +573,54 @@ func (d *Resolver) filterDisconnectedPeerAnswers(logger *log.Entry, question dns
 	return kept
 }
 
-// extractRecordIP returns the dotted-decimal / colon-hex IP carried by
-// an A or AAAA record, or "" for any other record type.
-func extractRecordIP(rr dns.RR) string {
+// warmLazyPeers triggers lazy-connection wake-up for the peers a resolved
+// answer points at and waits briefly for one to connect, so the caller's first
+// request doesn't race the connection establishment. Warm-up is scoped to
+// match-only (non-authoritative) zones — the synthesized private-service zones
+// and user-created zones whose records point at specific peers. The account's
+// peer zone is authoritative, so plain peer-name lookups never trigger warm-up;
+// otherwise resolving any peer's name would wake its idle connection, defeating
+// laziness mesh-wide. No-op when no activator is wired (lazy connections
+// disabled) or the answer carries no peer IPs.
+func (d *Resolver) warmLazyPeers(question dns.Question, records []dns.RR) {
+	d.mu.RLock()
+	activator := d.peerActivator
+	var nonAuth, found bool
+	if activator != nil {
+		nonAuth, found = d.findZone(question.Name)
+	}
+	d.mu.RUnlock()
+	if activator == nil || !found || !nonAuth {
+		return
+	}
+
+	var addrs []netip.Addr
+	for _, rr := range records {
+		if addr, ok := extractRecordAddr(rr); ok {
+			addrs = append(addrs, addr)
+		}
+	}
+	if len(addrs) == 0 {
+		return
+	}
+
+	ctx, cancel := context.WithTimeout(d.ctx, d.warmupTimeout)
+	defer cancel()
+	activator.ActivatePeersByIP(ctx, addrs)
+}
+
+// extractRecordAddr returns the IP address carried by an A or AAAA record.
+// ok is false for any other record type or a record with no address.
+func extractRecordAddr(rr dns.RR) (netip.Addr, bool) {
 	switch r := rr.(type) {
 	case *dns.A:
-		if r.A == nil {
-			return ""
-		}
-		return r.A.String()
+		addr, ok := netip.AddrFromSlice(r.A)
+		return addr.Unmap(), ok
 	case *dns.AAAA:
-		if r.AAAA == nil {
-			return ""
-		}
-		return r.AAAA.String()
+		addr, ok := netip.AddrFromSlice(r.AAAA)
+		return addr.Unmap(), ok
 	}
-	return ""
+	return netip.Addr{}, false
 }
 
 // Update replaces all zones and their records

client/internal/dns/local/local_test.go

@@ -37,8 +37,8 @@ type mockPeerConnectivity struct {
 	byIP map[string]struct{ known, connected bool }
 }
 
-func (m mockPeerConnectivity) IsConnectedByIP(ip string) (known, connected bool) {
-	v, ok := m.byIP[ip]
+func (m mockPeerConnectivity) IsConnectedByIP(ip netip.Addr) (known, connected bool) {
+	v, ok := m.byIP[ip.String()]
 	if !ok {
 		return false, false
 	}

client/internal/dns/local/warmup_test.go

@@ -0,0 +1,172 @@
+package local
+
+import (
+	"context"
+	"net"
+	"net/netip"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/miekg/dns"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/dns/test"
+	nbdns "github.com/netbirdio/netbird/dns"
+)
+
+// recordingActivator records the addresses it was asked to warm and returns
+// immediately, so ServeDNS is not blocked by the test.
+type recordingActivator struct {
+	mu     sync.Mutex
+	called bool
+	addrs  []netip.Addr
+}
+
+func (r *recordingActivator) ActivatePeersByIP(_ context.Context, addrs []netip.Addr) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	r.called = true
+	r.addrs = append(r.addrs, addrs...)
+}
+
+func serveA(t *testing.T, resolver *Resolver, name string) *dns.Msg {
+	t.Helper()
+	var resp *dns.Msg
+	w := &test.MockResponseWriter{WriteMsgFunc: func(m *dns.Msg) error { resp = m; return nil }}
+	resolver.ServeDNS(w, new(dns.Msg).SetQuestion(name, dns.TypeA))
+	return resp
+}
+
+// serviceZone registers rec in a match-only (non-authoritative) zone, the shape
+// the synthesized private-service zones arrive in.
+func serviceZone(t *testing.T, resolver *Resolver, zone string, records ...nbdns.SimpleRecord) {
+	t.Helper()
+	resolver.Update([]nbdns.CustomZone{{
+		Domain:           zone,
+		Records:          records,
+		NonAuthoritative: true,
+	}})
+}
+
+func TestLocalResolver_WarmsLazyPeerOnResolve(t *testing.T) {
+	rec := nbdns.SimpleRecord{Name: "svc.proxy.netbird.cloud.", Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.7"}
+	resolver := NewResolver()
+	serviceZone(t, resolver, "proxy.netbird.cloud", rec)
+
+	act := &recordingActivator{}
+	resolver.SetPeerActivator(act)
+
+	resp := serveA(t, resolver, rec.Name)
+	require.NotNil(t, resp, "resolver must answer")
+	require.NotEmpty(t, resp.Answer, "answer must carry the A record")
+
+	act.mu.Lock()
+	defer act.mu.Unlock()
+	assert.True(t, act.called, "activator must be invoked for a service-zone A answer")
+	assert.Contains(t, act.addrs, netip.MustParseAddr("100.64.0.7"), "activator must receive the answer's peer IP")
+}
+
+func TestLocalResolver_NoActivatorNoWarmup(t *testing.T) {
+	// With no activator wired the resolver behaves exactly as before.
+	rec := nbdns.SimpleRecord{Name: "svc.proxy.netbird.cloud.", Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.7"}
+	resolver := NewResolver()
+	serviceZone(t, resolver, "proxy.netbird.cloud", rec)
+
+	resp := serveA(t, resolver, rec.Name)
+	require.NotNil(t, resp, "resolver must still answer without an activator")
+	require.NotEmpty(t, resp.Answer, "answer must carry the A record")
+}
+
+func TestLocalResolver_NoWarmupForMissingRecord(t *testing.T) {
+	// A query that resolves to nothing must not invoke the activator (no IPs).
+	resolver := NewResolver()
+	serviceZone(t, resolver, "proxy.netbird.cloud",
+		nbdns.SimpleRecord{Name: "svc.proxy.netbird.cloud.", Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.7"})
+
+	act := &recordingActivator{}
+	resolver.SetPeerActivator(act)
+
+	serveA(t, resolver, "absent.proxy.netbird.cloud.")
+
+	act.mu.Lock()
+	defer act.mu.Unlock()
+	assert.False(t, act.called, "activator must not be invoked when there is no answer")
+}
+
+func TestLocalResolver_NoWarmupInAuthoritativeZone(t *testing.T) {
+	// The account's peer zone is authoritative; resolving a peer's name there
+	// must not wake its lazy connection — warm-up is scoped to match-only
+	// (non-authoritative) zones such as the synthesized private-service zones.
+	rec := nbdns.SimpleRecord{Name: "peer.netbird.cloud.", Type: 1, Class: nbdns.DefaultClass, TTL: 300, RData: "100.64.0.9"}
+	resolver := NewResolver()
+	resolver.Update([]nbdns.CustomZone{{
+		Domain:  "netbird.cloud",
+		Records: []nbdns.SimpleRecord{rec},
+	}})
+
+	act := &recordingActivator{}
+	resolver.SetPeerActivator(act)
+
+	resp := serveA(t, resolver, rec.Name)
+	require.NotNil(t, resp, "resolver must answer")
+	require.NotEmpty(t, resp.Answer, "answer must carry the A record")
+
+	act.mu.Lock()
+	defer act.mu.Unlock()
+	assert.False(t, act.called, "activator must not be invoked for authoritative-zone answers")
+}
+
+func TestLazyWarmupTimeoutFromEnv(t *testing.T) {
+	tests := []struct {
+		name   string
+		value  string
+		envSet bool
+		want   time.Duration
+	}{
+		{name: "unset uses default", want: defaultLazyWarmupTimeout},
+		{name: "valid overrides", value: "5s", envSet: true, want: 5 * time.Second},
+		{name: "invalid falls back", value: "not-a-duration", envSet: true, want: defaultLazyWarmupTimeout},
+		{name: "negative falls back", value: "-1s", envSet: true, want: defaultLazyWarmupTimeout},
+		{name: "zero falls back", value: "0s", envSet: true, want: defaultLazyWarmupTimeout},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if tt.envSet {
+				t.Setenv(envLazyWarmupTimeout, tt.value)
+			}
+			assert.Equal(t, tt.want, lazyWarmupTimeoutFromEnv())
+			assert.Equal(t, tt.want, NewResolver().warmupTimeout, "constructor must resolve the timeout once")
+		})
+	}
+}
+
+func TestExtractRecordAddr(t *testing.T) {
+	t.Run("A record yields unmapped v4", func(t *testing.T) {
+		// net.ParseIP returns the 16-byte v4-in-v6 form, the same shape
+		// miekg/dns stores after parsing an A record; the extracted address
+		// must compare equal to a plain v4 netip.Addr.
+		addr, ok := extractRecordAddr(&dns.A{A: net.ParseIP("100.64.0.7")})
+		require.True(t, ok)
+		assert.True(t, addr.Is4())
+		assert.Equal(t, netip.MustParseAddr("100.64.0.7"), addr)
+	})
+
+	t.Run("AAAA record yields v6", func(t *testing.T) {
+		addr, ok := extractRecordAddr(&dns.AAAA{AAAA: net.ParseIP("fd00::1")})
+		require.True(t, ok)
+		assert.Equal(t, netip.MustParseAddr("fd00::1"), addr)
+	})
+
+	t.Run("A record without address", func(t *testing.T) {
+		_, ok := extractRecordAddr(&dns.A{})
+		assert.False(t, ok)
+	})
+
+	t.Run("non-address record", func(t *testing.T) {
+		_, ok := extractRecordAddr(&dns.CNAME{Target: "target.netbird.cloud."})
+		assert.False(t, ok)
+	})
+}

client/internal/dns/mock_server.go

@@ -8,6 +8,7 @@ import (
 	"github.com/miekg/dns"
 
 	dnsconfig "github.com/netbirdio/netbird/client/internal/dns/config"
+	"github.com/netbirdio/netbird/client/internal/dns/local"
 	nbdns "github.com/netbirdio/netbird/dns"
 	"github.com/netbirdio/netbird/route"
 	"github.com/netbirdio/netbird/shared/management/domain"
@@ -92,6 +93,11 @@ func (m *MockServer) SetFirewall(Firewall) {
 	// Mock implementation - no-op
 }
 
+// SetPeerActivator mock implementation of SetPeerActivator from Server interface
+func (m *MockServer) SetPeerActivator(local.PeerActivator) {
+	// Mock implementation - no-op
+}
+
 // BeginBatch mock implementation of BeginBatch from Server interface
 func (m *MockServer) BeginBatch() {
 	// Mock implementation - no-op

client/internal/dns/server.go

@@ -82,6 +82,7 @@ type Server interface {
 	PopulateManagementDomain(mgmtURL *url.URL) error
 	SetRouteSources(selected, active func() route.HAMap)
 	SetFirewall(Firewall)
+	SetPeerActivator(local.PeerActivator)
 }
 
 type nsGroupsByDomain struct {
@@ -491,6 +492,13 @@ func (s *DefaultServer) SetFirewall(fw Firewall) {
 	}
 }
 
+// SetPeerActivator wires the DNS-time lazy-connection warm-up on the local
+// resolver. Injected after the connection manager exists (it does not at
+// DNS-server construction time). Pass nil to disable.
+func (s *DefaultServer) SetPeerActivator(a local.PeerActivator) {
+	s.localResolver.SetPeerActivator(a)
+}
+
 // Stop stops the server
 func (s *DefaultServer) Stop() {
 	s.ctxCancel()
@@ -1435,11 +1443,11 @@ type localPeerConnectivity struct {
 
 // IsConnectedByIP looks the IP up in the peerstore and surfaces both
 // the known and connected bits. Used by Resolver.filterDisconnectedPeerAnswers.
-func (l localPeerConnectivity) IsConnectedByIP(ip string) (known, connected bool) {
+func (l localPeerConnectivity) IsConnectedByIP(ip netip.Addr) (known, connected bool) {
 	if l.status == nil {
 		return false, false
 	}
-	state, ok := l.status.PeerStateByIP(ip)
+	state, ok := l.status.PeerStateByIP(ip.String())
 	if !ok {
 		return false, false
 	}

client/internal/dns_peer_activator.go

@@ -0,0 +1,76 @@
+package internal
+
+import (
+	"context"
+	"net/netip"
+	"time"
+
+	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/client/internal/peerstore"
+)
+
+const dnsActivationPollInterval = 50 * time.Millisecond
+
+// dnsPeerActivator wakes lazy-connection peers from the DNS resolution path. It
+// implements dns/local.PeerActivator. DNS queries run on their own goroutines,
+// so it only touches state that is safe for concurrent use — ConnMgr.ActivatePeer,
+// peerstore.Store and peer.Status — and never takes the engine's syncMsgMux,
+// keeping DNS resolution from contending with network-map processing.
+type dnsPeerActivator struct {
+	connMgr   *ConnMgr
+	peerStore *peerstore.Store
+	status    *peer.Status
+	// ctx is the engine's long-lived context. The connection dial is tied to it
+	// (not the per-query DNS wait budget) so a handshake that outlasts the wait
+	// still completes in the background rather than being cancelled at the deadline.
+	ctx context.Context
+}
+
+// ActivatePeersByIP triggers wake-up for the peer(s) owning addrs and waits
+// until one is connected or ctx (the per-query DNS wait budget) expires.
+// Activation itself is tied to the engine's long-lived context so the dial
+// survives a wait that times out. Unknown or already-connected addresses are
+// skipped, so the steady-state (warm) path adds no latency.
+func (a *dnsPeerActivator) ActivatePeersByIP(ctx context.Context, addrs []netip.Addr) {
+	if a == nil || a.connMgr == nil {
+		return
+	}
+
+	var pending []string
+	for _, addr := range addrs {
+		ip := addr.String()
+		st, ok := a.status.PeerStateByIP(ip)
+		if !ok || st.ConnStatus == peer.StatusConnected {
+			continue
+		}
+		conn, ok := a.peerStore.PeerConn(st.PubKey)
+		if !ok {
+			continue
+		}
+		a.connMgr.ActivatePeer(a.ctx, conn)
+		pending = append(pending, ip)
+	}
+
+	if len(pending) == 0 {
+		return
+	}
+	a.waitConnected(ctx, pending)
+}
+
+// waitConnected blocks until any of ips reports a connected peer or ctx expires.
+func (a *dnsPeerActivator) waitConnected(ctx context.Context, ips []string) {
+	ticker := time.NewTicker(dnsActivationPollInterval)
+	defer ticker.Stop()
+	for {
+		for _, ip := range ips {
+			if st, ok := a.status.PeerStateByIP(ip); ok && st.ConnStatus == peer.StatusConnected {
+				return
+			}
+		}
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+	}
+}

client/internal/dns_peer_activator_test.go

@@ -0,0 +1,129 @@
+package internal
+
+import (
+	"context"
+	"net/netip"
+	"testing"
+	"time"
+
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/client/internal/peer"
+	"github.com/netbirdio/netbird/client/internal/peerstore"
+)
+
+func newTestPeerConn(t *testing.T, key string) *peer.Conn {
+	t.Helper()
+	conn, err := peer.NewConn(peer.ConnConfig{
+		Key:      key,
+		LocalKey: "local",
+		WgConfig: peer.WgConfig{
+			AllowedIps: []netip.Prefix{netip.MustParsePrefix("100.64.0.1/32")},
+		},
+	}, peer.ServiceDependencies{})
+	require.NoError(t, err)
+	return conn
+}
+
+func newTestDNSPeerActivator(t *testing.T) (*dnsPeerActivator, *peer.Status, *peerstore.Store) {
+	t.Helper()
+	status := peer.NewRecorder("https://mgm")
+	store := peerstore.NewConnStore()
+	// ConnMgr without Start: the lazy manager is nil, so ActivatePeer is a
+	// no-op — these tests exercise the activator's skip/wait logic.
+	connMgr := NewConnMgr(&EngineConfig{}, status, store, nil)
+	return &dnsPeerActivator{
+		connMgr:   connMgr,
+		peerStore: store,
+		status:    status,
+		ctx:       context.Background(),
+	}, status, store
+}
+
+func TestDNSPeerActivator_NilSafe(t *testing.T) {
+	var a *dnsPeerActivator
+	a.ActivatePeersByIP(context.Background(), []netip.Addr{netip.MustParseAddr("100.64.0.1")})
+}
+
+// TestDNSPeerActivator_SkipsUnknownAndConnectedPeers verifies the steady-state
+// (warm) path adds no latency: already-connected and unknown addresses never
+// enter the wait loop.
+func TestDNSPeerActivator_SkipsUnknownAndConnectedPeers(t *testing.T) {
+	a, status, store := newTestDNSPeerActivator(t)
+
+	require.NoError(t, status.AddPeer("peerA", "a.netbird.cloud", "100.64.0.1", "fd00::1"))
+	require.NoError(t, status.UpdatePeerState(peer.State{PubKey: "peerA", ConnStatus: peer.StatusConnected}))
+	store.AddPeerConn("peerA", newTestPeerConn(t, "peerA"))
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	a.ActivatePeersByIP(ctx, []netip.Addr{
+		netip.MustParseAddr("100.64.0.1"),  // known, connected -> skipped
+		netip.MustParseAddr("fd00::1"),     // known via IPv6, connected -> skipped
+		netip.MustParseAddr("100.64.0.99"), // unknown -> skipped
+	})
+	require.Less(t, time.Since(start), time.Second, "no pending peer must mean no wait")
+}
+
+// TestDNSPeerActivator_WaitsForPendingPeerToConnect verifies the wait loop
+// returns as soon as a pending peer reports connected, well before the
+// per-query budget expires.
+func TestDNSPeerActivator_WaitsForPendingPeerToConnect(t *testing.T) {
+	a, status, store := newTestDNSPeerActivator(t)
+
+	require.NoError(t, status.AddPeer("peerA", "a.netbird.cloud", "100.64.0.1", ""))
+	store.AddPeerConn("peerA", newTestPeerConn(t, "peerA"))
+
+	go func() {
+		time.Sleep(150 * time.Millisecond)
+		_ = status.UpdatePeerState(peer.State{PubKey: "peerA", ConnStatus: peer.StatusConnected})
+	}()
+
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	a.ActivatePeersByIP(ctx, []netip.Addr{netip.MustParseAddr("100.64.0.1")})
+	elapsed := time.Since(start)
+
+	require.GreaterOrEqual(t, elapsed, 100*time.Millisecond, "must wait for the pending peer")
+	require.Less(t, elapsed, 5*time.Second, "must return on connect, not at the deadline")
+}
+
+// TestDNSPeerActivator_ReturnsAtBudgetWhenPeerStaysIdle verifies a peer that
+// never connects releases the DNS response at the per-query budget instead of
+// blocking it indefinitely.
+func TestDNSPeerActivator_ReturnsAtBudgetWhenPeerStaysIdle(t *testing.T) {
+	a, status, store := newTestDNSPeerActivator(t)
+
+	require.NoError(t, status.AddPeer("peerA", "a.netbird.cloud", "100.64.0.1", ""))
+	store.AddPeerConn("peerA", newTestPeerConn(t, "peerA"))
+
+	ctx, cancel := context.WithTimeout(context.Background(), 300*time.Millisecond)
+	defer cancel()
+
+	start := time.Now()
+	a.ActivatePeersByIP(ctx, []netip.Addr{netip.MustParseAddr("100.64.0.1")})
+	elapsed := time.Since(start)
+
+	require.GreaterOrEqual(t, elapsed, 250*time.Millisecond, "must wait out the budget for a pending peer")
+	require.Less(t, elapsed, 5*time.Second, "must not block past the budget")
+}
+
+// TestDNSPeerActivator_NoWaitWithoutPeerConn verifies a known-but-idle peer
+// with no connection object in the store is not waited on: there is nothing to
+// activate, so waiting could only ever time out.
+func TestDNSPeerActivator_NoWaitWithoutPeerConn(t *testing.T) {
+	a, status, _ := newTestDNSPeerActivator(t)
+
+	require.NoError(t, status.AddPeer("peerA", "a.netbird.cloud", "100.64.0.1", ""))
+
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+
+	start := time.Now()
+	a.ActivatePeersByIP(ctx, []netip.Addr{netip.MustParseAddr("100.64.0.1")})
+	require.Less(t, time.Since(start), time.Second, "peer without a conn must not be waited on")
+}

client/internal/engine.go

@@ -654,6 +654,16 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 	e.connMgr = NewConnMgr(e.config, e.statusRecorder, e.peerStore, wgIface)
 	e.connMgr.Start(e.ctx)
 
+	// Wire DNS-time lazy-connection warm-up now that the connection manager
+	// exists (it does not at DNS-server construction time). A DNS answer that
+	// points at an idle peer then wakes it before the client's first request.
+	e.dnsServer.SetPeerActivator(&dnsPeerActivator{
+		connMgr:   e.connMgr,
+		peerStore: e.peerStore,
+		status:    e.statusRecorder,
+		ctx:       e.ctx,
+	})
+
 	e.srWatcher = guard.NewSRWatcher(e.signal, e.relayManager, e.mobileDep.IFaceDiscover, iceCfg)
 	e.srWatcher.Start(peer.IsForceRelayed())
 

client/internal/routemanager/dynamic/route.go

@@ -185,7 +185,7 @@ func (r *Route) startResolver(ctx context.Context) {
 }
 
 func (r *Route) update(ctx context.Context) error {
-	resolved, err := r.resolveDomains()
+	resolved, err := r.resolveDomains(ctx)
 	if err != nil {
 		if len(resolved) == 0 {
 			return fmt.Errorf("resolve domains: %w", err)
@@ -199,9 +199,9 @@ func (r *Route) update(ctx context.Context) error {
 	return nil
 }
 
-func (r *Route) resolveDomains() (domainMap, error) {
+func (r *Route) resolveDomains(ctx context.Context) (domainMap, error) {
 	results := make(chan resolveResult)
-	go r.resolve(results)
+	go r.resolve(ctx, results)
 
 	resolved := domainMap{}
 	var merr *multierror.Error
@@ -217,7 +217,7 @@ func (r *Route) resolveDomains() (domainMap, error) {
 	return resolved, nberrors.FormatErrorOrNil(merr)
 }
 
-func (r *Route) resolve(results chan resolveResult) {
+func (r *Route) resolve(ctx context.Context, results chan resolveResult) {
 	var wg sync.WaitGroup
 
 	for _, d := range r.route.Domains {
@@ -225,10 +225,10 @@ func (r *Route) resolve(results chan resolveResult) {
 		go func(domain domain.Domain) {
 			defer wg.Done()
 
-			ips, err := r.getIPsFromResolver(domain)
+			ips, err := r.getIPsFromResolver(ctx, domain)
 			if err != nil {
 				log.Tracef("Failed to resolve domain %s with private resolver: %v", domain.SafeString(), err)
-				ips, err = net.LookupIP(domain.PunycodeString())
+				ips, err = lookupHostIPs(ctx, domain)
 				if err != nil {
 					results <- resolveResult{domain: domain, err: fmt.Errorf("resolve d %s: %w", domain.SafeString(), err)}
 					return
@@ -364,6 +364,20 @@ func determinePrefixChanges(oldPrefixes, newPrefixes []netip.Prefix) (toAdd, toR
 	return
 }
 
+// lookupHostIPs resolves d via the system resolver, honoring ctx cancellation.
+func lookupHostIPs(ctx context.Context, d domain.Domain) ([]net.IP, error) {
+	addrs, err := net.DefaultResolver.LookupIPAddr(ctx, d.PunycodeString())
+	if err != nil {
+		return nil, err
+	}
+
+	ips := make([]net.IP, 0, len(addrs))
+	for _, addr := range addrs {
+		ips = append(ips, addr.IP)
+	}
+	return ips, nil
+}
+
 func combinePrefixes(oldPrefixes, removedPrefixes, addedPrefixes []netip.Prefix) []netip.Prefix {
 	prefixSet := make(map[netip.Prefix]struct{})
 	for _, prefix := range oldPrefixes {

client/internal/routemanager/dynamic/route_generic.go

@@ -3,11 +3,12 @@
 package dynamic
 
 import (
+	"context"
 	"net"
 
 	"github.com/netbirdio/netbird/shared/management/domain"
 )
 
-func (r *Route) getIPsFromResolver(domain domain.Domain) ([]net.IP, error) {
-	return net.LookupIP(domain.PunycodeString())
+func (r *Route) getIPsFromResolver(ctx context.Context, domain domain.Domain) ([]net.IP, error) {
+	return lookupHostIPs(ctx, domain)
 }

client/internal/routemanager/dynamic/route_ios.go

@@ -3,6 +3,7 @@
 package dynamic
 
 import (
+	"context"
 	"fmt"
 	"net"
 	"time"
@@ -16,7 +17,7 @@ import (
 
 const dialTimeout = 10 * time.Second
 
-func (r *Route) getIPsFromResolver(domain domain.Domain) ([]net.IP, error) {
+func (r *Route) getIPsFromResolver(ctx context.Context, domain domain.Domain) ([]net.IP, error) {
 	privateClient, err := nbdns.GetClientPrivate(r.wgInterface, r.resolverAddr.Addr(), dialTimeout)
 	if err != nil {
 		return nil, fmt.Errorf("error while creating private client: %s", err)
@@ -32,7 +33,7 @@ func (r *Route) getIPsFromResolver(domain domain.Domain) ([]net.IP, error) {
 		msg := new(dns.Msg)
 		msg.SetQuestion(fqdn, qtype)
 
-		response, _, err := nbdns.ExchangeWithFallback(nil, privateClient, msg, r.resolverAddr.String())
+		response, _, err := nbdns.ExchangeWithFallback(ctx, privateClient, msg, r.resolverAddr.String())
 		if err != nil {
 			if queryErr == nil {
 				queryErr = fmt.Errorf("DNS query for %s (type %d) after %s: %w", domain.SafeString(), qtype, time.Since(startTime), err)

client/internal/sleep/detector_notsupported.go

@@ -1,4 +1,4 @@
-//go:build (!darwin && !windows) || ios
+//go:build !darwin || ios
 
 package sleep
 

client/internal/sleep/detector_windows.go

@@ -1,255 +0,0 @@
-//go:build windows
-
-package sleep
-
-import (
-	"fmt"
-	"sync"
-	"time"
-	"unsafe"
-
-	log "github.com/sirupsen/logrus"
-	"golang.org/x/sys/windows"
-)
-
-// Power broadcast event types delivered to the notification callback.
-// https://learn.microsoft.com/en-us/windows/win32/power/power-management-events
-const (
-	pbtAPMSuspend         uintptr = 0x0004
-	pbtAPMResumeSuspend   uintptr = 0x0007
-	pbtAPMResumeAutomatic uintptr = 0x0012
-	deviceNotifyCallback  uintptr = 0x00000002
-)
-
-var (
-	// The callback flavor of these functions is exported by user32.dll, not
-	// powrprof.dll (which exports the differently-shaped Power* variants).
-	user32 = windows.NewLazySystemDLL("user32.dll")
-
-	// https://learn.microsoft.com/en-us/windows/win32/api/winuser/nf-winuser-registersuspendresumenotification
-	registerSuspendResumeNotification = user32.NewProc("RegisterSuspendResumeNotification")
-	// https://learn.microsoft.com/en-us/windows/win32/api/winuser/nf-winuser-unregistersuspendresumenotification
-	unregisterSuspendResumeNotification = user32.NewProc("UnregisterSuspendResumeNotification")
-
-	libInitOnce sync.Once
-	libInitErr  error
-
-	// callbackThunk is the single C-callable trampoline registered with the OS.
-	// windows.NewCallback slots are a finite, non-reclaimable resource, so a
-	// single thunk dispatches to the Detector identified by the callback Context.
-	callbackThunk uintptr
-
-	// registry maps the Context value handed to the OS callback back to the
-	// Detector that registered it, mirroring darwin's serviceRegistry.
-	registry   = make(map[int]*Detector)
-	registryMu sync.Mutex
-	nextHandle int
-
-	// lifecycleMu serializes Register/Deregister so concurrent lifecycle calls
-	// can't race on the shared registry or the OS registration handle.
-	lifecycleMu sync.Mutex
-)
-
-// deviceNotifySubscribeParameters is DEVICE_NOTIFY_SUBSCRIBE_PARAMETERS.
-// https://learn.microsoft.com/en-us/windows/win32/api/winuser/ns-winuser-device_notify_subscribe_parameters
-type deviceNotifySubscribeParameters struct {
-	Callback uintptr
-	Context  uintptr
-}
-
-// Detector delivers sleep and wake events to a registered callback.
-type Detector struct {
-	callback func(event EventType)
-	done     chan struct{}
-
-	// handle keys this detector in the package registry and is passed to the OS
-	// as the callback Context. Zero means the detector is not registered.
-	handle int
-	// hPowerNotify is the HPOWERNOTIFY returned by RegisterSuspendResumeNotification.
-	hPowerNotify uintptr
-	// params is kept alive for the lifetime of the registration so the OS never
-	// dereferences freed memory.
-	params deviceNotifySubscribeParameters
-}
-
-// NewDetector resolves powrprof.dll symbols and returns a Detector.
-func NewDetector() (*Detector, error) {
-	if err := initLibs(); err != nil {
-		return nil, err
-	}
-	return &Detector{}, nil
-}
-
-func initLibs() error {
-	libInitOnce.Do(func() {
-		if err := registerSuspendResumeNotification.Find(); err != nil {
-			libInitErr = fmt.Errorf("resolve RegisterSuspendResumeNotification: %w", err)
-			return
-		}
-		if err := unregisterSuspendResumeNotification.Find(); err != nil {
-			libInitErr = fmt.Errorf("resolve UnregisterSuspendResumeNotification: %w", err)
-			return
-		}
-		callbackThunk = windows.NewCallback(powerCallback)
-	})
-	return libInitErr
-}
-
-// Register installs callback for power events and subscribes to suspend/resume
-// notifications via powrprof.dll.
-func (d *Detector) Register(callback func(event EventType)) error {
-	lifecycleMu.Lock()
-	defer lifecycleMu.Unlock()
-
-	registryMu.Lock()
-	if d.handle != 0 {
-		registryMu.Unlock()
-		return fmt.Errorf("detector service already registered")
-	}
-	d.callback = callback
-	d.done = make(chan struct{})
-	nextHandle++
-	handle := nextHandle
-	d.handle = handle
-	registry[handle] = d
-	registryMu.Unlock()
-
-	d.params = deviceNotifySubscribeParameters{
-		Callback: callbackThunk,
-		Context:  uintptr(handle),
-	}
-
-	ret, _, callErr := registerSuspendResumeNotification.Call(
-		uintptr(unsafe.Pointer(&d.params)),
-		deviceNotifyCallback,
-	)
-	if ret == 0 {
-		registryMu.Lock()
-		delete(registry, handle)
-		close(d.done)
-		d.done = nil
-		d.handle = 0
-		registryMu.Unlock()
-		return fmt.Errorf("RegisterSuspendResumeNotification failed: %w", callErr)
-	}
-	d.hPowerNotify = ret
-
-	log.Info("sleep detection service started on Windows")
-	return nil
-}
-
-// Deregister unsubscribes from power notifications and removes the detector.
-func (d *Detector) Deregister() error {
-	lifecycleMu.Lock()
-	defer lifecycleMu.Unlock()
-
-	registryMu.Lock()
-	if d.handle == 0 {
-		registryMu.Unlock()
-		return nil
-	}
-	handle := d.handle
-	hPowerNotify := d.hPowerNotify
-	done := d.done
-	registryMu.Unlock()
-
-	log.Info("sleep detection service stopping (deregister)")
-
-	// Unregister the OS subscription first. If it fails, leave handle and
-	// hPowerNotify intact so a later call can retry the cleanup.
-	if hPowerNotify != 0 {
-		ret, _, callErr := unregisterSuspendResumeNotification.Call(hPowerNotify)
-		if ret == 0 {
-			return fmt.Errorf("UnregisterSuspendResumeNotification failed: %w", callErr)
-		}
-	}
-
-	registryMu.Lock()
-	close(done)
-	delete(registry, handle)
-	d.handle = 0
-	d.hPowerNotify = 0
-	registryMu.Unlock()
-
-	return nil
-}
-
-func (d *Detector) triggerCallback(event EventType, cb func(event EventType), done <-chan struct{}) {
-	if cb == nil || done == nil {
-		return
-	}
-
-	select {
-	case <-done:
-		return
-	default:
-	}
-
-	doneChan := make(chan struct{})
-	// The OS invokes this callback synchronously on the suspend path, so run the
-	// teardown inline with a bounded budget (mirroring the macOS detector) so
-	// Down completes before the machine suspends without blocking indefinitely.
-	timeout := time.NewTimer(20 * time.Second)
-	defer timeout.Stop()
-
-	go func() {
-		defer close(doneChan)
-		defer func() {
-			if r := recover(); r != nil {
-				log.Errorf("panic in sleep callback: %v", r)
-			}
-		}()
-		log.Info("sleep detection event fired")
-		cb(event)
-	}()
-
-	select {
-	case <-doneChan:
-	case <-done:
-	case <-timeout.C:
-		log.Warn("sleep callback timed out")
-	}
-}
-
-// powerCallback is the DEVICE_NOTIFY_CALLBACK_ROUTINE trampoline, invoked by the
-// OS on a system thread. A Go panic crossing the syscall boundary has undefined
-// behavior, so contain it here. It must return ERROR_SUCCESS (0).
-func powerCallback(context uintptr, msgType uintptr, setting uintptr) uintptr {
-	defer func() {
-		if r := recover(); r != nil {
-			log.Errorf("panic in sleep powerCallback: %v", r)
-		}
-	}()
-
-	var event EventType
-	switch msgType {
-	case pbtAPMSuspend:
-		event = EventTypeSleep
-	case pbtAPMResumeAutomatic, pbtAPMResumeSuspend:
-		event = EventTypeWakeUp
-	default:
-		return 0
-	}
-
-	dispatchEvent(int(context), event)
-	return 0
-}
-
-func dispatchEvent(handle int, event EventType) {
-	registryMu.Lock()
-	d := registry[handle]
-	var (
-		cb   func(event EventType)
-		done <-chan struct{}
-	)
-	if d != nil {
-		cb = d.callback
-		done = d.done
-	}
-	registryMu.Unlock()
-
-	if d == nil {
-		return
-	}
-	d.triggerCallback(event, cb, done)
-}

client/internal/sleep/detector_windows_test.go

@@ -1,83 +0,0 @@
-//go:build windows
-
-package sleep
-
-import (
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// registerFake inserts a detector into the registry without touching the OS,
-// so dispatch routing can be exercised in isolation. It returns the handle and
-// a cleanup that removes the entry.
-func registerFake(t *testing.T, cb func(EventType)) (int, func()) {
-	t.Helper()
-
-	registryMu.Lock()
-	nextHandle++
-	handle := nextHandle
-	d := &Detector{callback: cb, done: make(chan struct{}), handle: handle}
-	registry[handle] = d
-	registryMu.Unlock()
-
-	return handle, func() {
-		registryMu.Lock()
-		delete(registry, handle)
-		registryMu.Unlock()
-	}
-}
-
-func TestPowerCallback_MapsMessageTypes(t *testing.T) {
-	tests := []struct {
-		name    string
-		msgType uintptr
-		want    EventType
-		fires   bool
-	}{
-		{"suspend", pbtAPMSuspend, EventTypeSleep, true},
-		{"resume automatic", pbtAPMResumeAutomatic, EventTypeWakeUp, true},
-		{"resume suspend", pbtAPMResumeSuspend, EventTypeWakeUp, true},
-		{"unknown", 0x9999, EventTypeUnknown, false},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			got := make(chan EventType, 1)
-			handle, cleanup := registerFake(t, func(e EventType) { got <- e })
-			defer cleanup()
-
-			ret := powerCallback(uintptr(handle), tt.msgType, 0)
-			require.Equal(t, uintptr(0), ret, "callback must return ERROR_SUCCESS")
-
-			if !tt.fires {
-				assert.Empty(t, got, "no event should fire for unhandled message type")
-				return
-			}
-			select {
-			case e := <-got:
-				assert.Equal(t, tt.want, e, "mapped event type should match")
-			default:
-				t.Fatal("expected callback to fire")
-			}
-		})
-	}
-}
-
-func TestDispatchEvent_UnknownHandleNoPanic(t *testing.T) {
-	require.NotPanics(t, func() {
-		dispatchEvent(-1, EventTypeSleep)
-	}, "dispatch for an unregistered handle must be a no-op")
-}
-
-func TestTriggerCallback_SkipsAfterDone(t *testing.T) {
-	done := make(chan struct{})
-	close(done)
-
-	fired := false
-	d := &Detector{}
-	d.triggerCallback(EventTypeSleep, func(EventType) { fired = true }, done)
-
-	assert.False(t, fired, "callback must not run once the detector is done")
-}

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/server/server.go

@@ -1081,7 +1081,10 @@ func (s *Server) Down(ctx context.Context, _ *proto.DownRequest) (*proto.DownRes
 
 	if err := s.cleanupConnection(); err != nil {
 		s.mutex.Unlock()
-		// todo review to update the status in case any type of error
+		if errors.Is(err, ErrServiceNotUp) {
+			log.Debugf("Down called while service not up: %v", err)
+			return nil, err
+		}
 		log.Errorf("failed to shut down properly: %v", err)
 		return nil, err
 	}
@@ -1154,7 +1157,7 @@ func (s *Server) cleanupConnection() error {
 	// making the run loop the sole owner of engine shutdown.
 	if engine != nil {
 		if err := engine.Stop(); err != nil {
-			return err
+			log.Errorf("failed to stop engine during cleanup: %v", err)
 		}
 	}
 

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/ui/frontend/src/components/empty-state/DaemonOutdatedOverlay.tsx

@@ -1,10 +1,13 @@
+import { useEffect, useState } from "react";
 import { useTranslation } from "react-i18next";
 import { AlertTriangleIcon, DownloadIcon } from "lucide-react";
 import { Browser } from "@wailsio/runtime";
+import { Version } from "@bindings/services";
 import { Button } from "@/components/buttons/Button";
 import { useStatus } from "@/contexts/StatusContext.tsx";
 
 const RELEASES_URL = "https://github.com/netbirdio/netbird/releases/latest";
+const RC_RELEASES_URL = "https://pkgs.netbird.io/releases/rc";
 
 function openUrl(url: string) {
     Browser.OpenURL(url).catch(() => globalThis.open(url, "_blank"));
@@ -12,7 +15,26 @@ function openUrl(url: string) {
 
 export const DaemonOutdatedOverlay = () => {
     const { t } = useTranslation();
-    const { isDaemonOutdated } = useStatus();
+    const { status, isDaemonOutdated } = useStatus();
+
+    const [guiVersion, setGuiVersion] = useState<string>("-");
+    const clientVersion = status?.daemonVersion ?? "—";
+
+    const isRc = /-rc/i.test(guiVersion) || /-rc/i.test(clientVersion);
+    const downloadUrl = isRc ? RC_RELEASES_URL : RELEASES_URL;
+
+    useEffect(() => {
+        if (!isDaemonOutdated) return;
+        let cancelled = false;
+        Version.GUI()
+            .then((v) => {
+                if (!cancelled) setGuiVersion(v);
+            })
+            .catch((err) => console.error("[DaemonOutdatedOverlay] GUI version error", err));
+        return () => {
+            cancelled = true;
+        };
+    }, [isDaemonOutdated]);
 
     if (!isDaemonOutdated) return null;
 
@@ -38,10 +60,37 @@ export const DaemonOutdatedOverlay = () => {
                     <p className={"text-sm text-nb-gray-300"}>{t("daemon.outdated.description")}</p>
                 </div>
 
+                <div className={"flex flex-col items-center gap-0.5 text-center"}>
+                    <p className={"text-sm font-semibold text-nb-gray-100"}>
+                        {clientVersion === "development" ? (
+                            <span>
+                                {t("settings.about.clientName")}{" "}
+                                <span className={"font-mono text-yellow-400"}>
+                                    {t("settings.about.development")}
+                                </span>
+                            </span>
+                        ) : (
+                            t("settings.about.client", { version: clientVersion })
+                        )}
+                    </p>
+                    <p className={"text-sm font-medium text-nb-gray-250"}>
+                        {guiVersion === "development" ? (
+                            <span>
+                                {t("settings.about.guiName")}{" "}
+                                <span className={"font-mono text-yellow-400"}>
+                                    {t("settings.about.development")}
+                                </span>
+                            </span>
+                        ) : (
+                            t("settings.about.gui", { version: guiVersion })
+                        )}
+                    </p>
+                </div>
+
                 <div className={"wails-no-draggable"}>
-                    <Button variant={"primary"} size={"xs"} onClick={() => openUrl(RELEASES_URL)}>
+                    <Button variant={"primary"} size={"xs"} onClick={() => openUrl(downloadUrl)}>
                         <DownloadIcon size={14} />
-                        {t("update.card.getInstaller")}
+                        {t("daemon.outdated.download")}
                     </Button>
                 </div>
             </div>

client/ui/frontend/src/contexts/ProfileContext.tsx

@@ -28,6 +28,7 @@ type ProfileContextValue = {
     loaded: boolean;
     refresh: () => Promise<void>;
     switchProfile: (id: string) => Promise<void>;
+    switchProfileNoConnect: (id: string) => Promise<void>;
     addProfile: (name: string) => Promise<string>;
     removeProfile: (id: string) => Promise<void>;
     renameProfile: (id: string, newName: string) => Promise<void>;
@@ -112,6 +113,16 @@ export const ProfileProvider = ({ children }: { children: ReactNode }) => {
         [username, refresh],
     );
 
+    // Manage-profiles variant: switches without connecting, so the user can
+    // still adjust the management URL before bringing the connection up.
+    const switchProfileNoConnect = useCallback(
+        async (id: string) => {
+            await ProfileSwitcher.SwitchActiveNoConnect({ profileName: id, username });
+            await refresh();
+        },
+        [username, refresh],
+    );
+
     // addProfile creates a profile by display name and returns the
     // daemon-generated ID, so the caller can immediately address it by ID.
     const addProfile = useCallback(
@@ -158,6 +169,7 @@ export const ProfileProvider = ({ children }: { children: ReactNode }) => {
             loaded,
             refresh,
             switchProfile,
+            switchProfileNoConnect,
             addProfile,
             removeProfile,
             renameProfile,
@@ -171,6 +183,7 @@ export const ProfileProvider = ({ children }: { children: ReactNode }) => {
             loaded,
             refresh,
             switchProfile,
+            switchProfileNoConnect,
             addProfile,
             removeProfile,
             renameProfile,

client/ui/frontend/src/modules/profiles/ProfilesTab.tsx

@@ -45,7 +45,7 @@ export function ProfilesTab() {
         activeProfileId,
         loaded,
         username,
-        switchProfile,
+        switchProfileNoConnect,
         addProfile,
         removeProfile,
         renameProfile,
@@ -100,7 +100,7 @@ export function ProfilesTab() {
             confirmLabel: t("profile.switch.confirm"),
         });
         if (!ok) return;
-        await guarded(i18next.t("profile.error.switchTitle"), () => switchProfile(id));
+        await guarded(i18next.t("profile.error.switchTitle"), () => switchProfileNoConnect(id));
     };
 
     const handleDeregister = async (id: string, name: string) => {
@@ -129,14 +129,13 @@ export function ProfilesTab() {
         await guarded(i18next.t("profile.error.createTitle"), async () => {
             const id = await addProfile(name);
             // SetConfig is keyed by the new profile's ID, so it writes the
-            // not-yet-active profile. Write before switching so any reconnect
-            // targets the right deployment.
+            // not-yet-active profile before the switch makes it current.
             if (!isNetbirdCloud(managementUrl)) {
                 await SettingsSvc.SetConfig(
                     new SetConfigParams({ profileName: id, username, managementUrl }),
                 );
             }
-            await switchProfile(id);
+            await switchProfileNoConnect(id);
         });
     };
 

client/ui/frontend/src/modules/session/SessionExpirationDialog.tsx

@@ -73,6 +73,13 @@ export default function SessionExpirationDialog() {
 
         let offCancel: (() => void) | undefined;
 
+        // Return the dialog to its interactive state and dismiss the browser popup
+        const resetDialog = () => {
+            offCancel?.();
+            WindowManager.CloseBrowserLogin().catch(console.error);
+            setBusy(false);
+        };
+
         try {
             const start = await Session.RequestExtend({ hint: "" });
             const uri = start.verificationUriComplete || start.verificationUri;
@@ -105,25 +112,22 @@ export default function SessionExpirationDialog() {
             if (outcome.kind === "cancel") {
                 waitPromise.cancel?.();
                 waitPromise.catch(() => {});
+                resetDialog();
                 return;
             }
 
             // Another surface owns this flow; keep the dialog open to retry.
             if (outcome.result.preempted) {
+                resetDialog();
                 return;
             }
-
-            // Close before the popup so the restore can't flash this window back.
-            WindowManager.CloseSessionExpiration().catch(console.error);
+            WindowManager.CloseRenewFlow().catch(console.error);
         } catch (e) {
+            resetDialog();
             await errorDialog({
                 Title: t("sessionExpiration.extendFailedTitle"),
                 Message: formatErrorMessage(e),
             });
-        } finally {
-            offCancel?.();
-            WindowManager.CloseBrowserLogin().catch(console.error);
-            setBusy(false);
         }
     }, [busy, t]);
 
@@ -139,12 +143,11 @@ export default function SessionExpirationDialog() {
             });
             WindowManager.CloseSessionExpiration().catch(console.error);
         } catch (e) {
+            setBusy(false);
             await errorDialog({
                 Title: t("sessionExpiration.logoutFailedTitle"),
                 Message: formatErrorMessage(e),
             });
-        } finally {
-            setBusy(false);
         }
     }, [busy, t]);
 

client/ui/frontend/src/modules/welcome/WelcomeStepTray.tsx

@@ -22,6 +22,9 @@ type WelcomeStepTrayProps = {
 export function WelcomeStepTray({ onContinue }: Readonly<WelcomeStepTrayProps>) {
     const { t } = useTranslation();
     const trayScreenshot = trayScreenshotForOS();
+    // macOS has no tray — the icon sits in the menu bar, so the copy says so.
+    const titleKey = isMacOS() ? "welcome.titleMac" : "welcome.title";
+    const descriptionKey = isMacOS() ? "welcome.descriptionMac" : "welcome.description";
 
     return (
         <>
@@ -36,9 +39,9 @@ export function WelcomeStepTray({ onContinue }: Readonly<WelcomeStepTrayProps>)
 
             <div className={"flex w-full flex-col gap-1"}>
                 <DialogHeading id={"nb-welcome-title"} align={"left"}>
-                    {t("welcome.title")}
+                    {t(titleKey)}
                 </DialogHeading>
-                <DialogDescription align={"left"}>{t("welcome.description")}</DialogDescription>
+                <DialogDescription align={"left"}>{t(descriptionKey)}</DialogDescription>
             </div>
 
             <DialogActions>

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/i18n/locales/de/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Suchen Sie NetBird in der Taskleiste"
     },
+    "welcome.titleMac": {
+        "message": "Suchen Sie NetBird in der Menüleiste"
+    },
     "welcome.description": {
         "message": "NetBird läuft in Ihrer Taskleiste. Klicken Sie auf das Symbol, um sich zu verbinden, Profile zu wechseln oder die Einstellungen zu öffnen."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird läuft in Ihrer Menüleiste. Klicken Sie auf das Symbol, um sich zu verbinden, Profile zu wechseln oder die Einstellungen zu öffnen."
+    },
     "welcome.continue": {
         "message": "Weiter"
     },
@@ -1293,10 +1299,13 @@
         "message": "Dokumentation"
     },
     "daemon.outdated.title": {
-        "message": "NetBird-Dienst ist veraltet"
+        "message": "NetBird Client ist veraltet"
     },
     "daemon.outdated.description": {
-        "message": "Aktualisieren Sie den NetBird-Dienst, um diese App zu verwenden."
+        "message": "Die neue GUI ist nicht mit Ihrem älteren Client kompatibel. Aktualisieren Sie Ihren Client, um die neue Anwendung zu verwenden."
+    },
+    "daemon.outdated.download": {
+        "message": "Neueste Version herunterladen"
     },
     "error.jwt_clock_skew": {
         "message": "Anmeldung fehlgeschlagen: Die Uhr dieses Geräts ist nicht mit dem Server synchron. Bitte synchronisieren Sie die Systemuhr und versuchen Sie es erneut."

client/ui/i18n/locales/en/common.json

@@ -1377,11 +1377,19 @@
     },
     "welcome.title": {
         "message": "Look for NetBird in your tray",
-        "description": "Heading on the first onboarding step, pointing the user to the tray icon. 'tray' = system tray / menu bar."
+        "description": "Heading on the first onboarding step, pointing the user to the tray icon. Shown on Windows and Linux; macOS uses welcome.titleMac."
+    },
+    "welcome.titleMac": {
+        "message": "Look for NetBird in your menu bar",
+        "description": "Heading on the first onboarding step on macOS, pointing the user to the menu bar icon. Use your language's Apple term for the macOS menu bar."
     },
     "welcome.description": {
         "message": "NetBird lives in your tray. Click the icon to connect, switch profiles, or open settings.",
-        "description": "Body of the first onboarding step explaining the tray icon."
+        "description": "Body of the first onboarding step explaining the tray icon. Shown on Windows and Linux; macOS uses welcome.descriptionMac."
+    },
+    "welcome.descriptionMac": {
+        "message": "NetBird lives in your menu bar. Click the icon to connect, switch profiles, or open settings.",
+        "description": "Body of the first onboarding step on macOS explaining the menu bar icon. Use your language's Apple term for the macOS menu bar."
     },
     "welcome.continue": {
         "message": "Continue",
@@ -1724,12 +1732,16 @@
         "description": "Documentation link on the daemon-unavailable overlay."
     },
     "daemon.outdated.title": {
-        "message": "NetBird Service Is Outdated",
-        "description": "Title of the overlay shown when the NetBird background service is too old to drive this UI."
+        "message": "NetBird Client Is Outdated",
+        "description": "Title of the overlay shown when the NetBird client (daemon) is too old to drive this UI."
     },
     "daemon.outdated.description": {
-        "message": "Update the NetBird service to use this app.",
-        "description": "Body of the daemon-outdated overlay telling the user to upgrade the service."
+        "message": "The new GUI isn't compatible with the older NetBird client. Update your client to use the new application.",
+        "description": "Body of the daemon-outdated overlay explaining that the GUI is newer than the client and the client must be updated."
+    },
+    "daemon.outdated.download": {
+        "message": "Download Latest",
+        "description": "Button on the daemon-outdated overlay that opens the download page for the latest release."
     },
     "error.jwt_clock_skew": {
         "message": "Sign-in failed: this device's clock is out of sync with the server. Please sync your system clock and try again.",

client/ui/i18n/locales/es/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Busque NetBird en su bandeja del sistema"
     },
+    "welcome.titleMac": {
+        "message": "Busque NetBird en su barra de menús"
+    },
     "welcome.description": {
         "message": "NetBird reside en su bandeja del sistema. Haga clic en el icono para conectarse, cambiar de perfil o abrir la configuración."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird reside en su barra de menús. Haga clic en el icono para conectarse, cambiar de perfil o abrir la configuración."
+    },
     "welcome.continue": {
         "message": "Continuar"
     },
@@ -1293,10 +1299,13 @@
         "message": "Documentación"
     },
     "daemon.outdated.title": {
-        "message": "El servicio de NetBird está desactualizado"
+        "message": "NetBird Client está desactualizado"
     },
     "daemon.outdated.description": {
-        "message": "Actualice el servicio de NetBird para usar esta aplicación."
+        "message": "La nueva GUI no es compatible con su cliente anterior. Actualice su cliente para usar la nueva aplicación."
+    },
+    "daemon.outdated.download": {
+        "message": "Descargar la última versión"
     },
     "error.jwt_clock_skew": {
         "message": "Error al iniciar sesión: el reloj de este dispositivo no está sincronizado con el servidor. Sincronice el reloj del sistema e inténtelo de nuevo."

client/ui/i18n/locales/fr/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Cherchez NetBird dans votre barre d’état système"
     },
+    "welcome.titleMac": {
+        "message": "Cherchez NetBird dans votre barre des menus"
+    },
     "welcome.description": {
         "message": "NetBird se trouve dans votre barre d’état système. Cliquez sur l’icône pour vous connecter, changer de profil ou ouvrir les paramètres."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird se trouve dans votre barre des menus. Cliquez sur l’icône pour vous connecter, changer de profil ou ouvrir les paramètres."
+    },
     "welcome.continue": {
         "message": "Continuer"
     },
@@ -1293,10 +1299,13 @@
         "message": "Documentation"
     },
     "daemon.outdated.title": {
-        "message": "Le service NetBird est obsolète"
+        "message": "Le Client NetBird est obsolète"
     },
     "daemon.outdated.description": {
-        "message": "Mettez à jour le service NetBird pour utiliser cette application."
+        "message": "La nouvelle GUI n'est pas compatible avec votre ancien client. Mettez à jour votre client pour utiliser la nouvelle application."
+    },
+    "daemon.outdated.download": {
+        "message": "Télécharger la dernière version"
     },
     "error.jwt_clock_skew": {
         "message": "Échec de la connexion : l’horloge de cet appareil n’est pas synchronisée avec le serveur. Veuillez synchroniser l’horloge de votre système et réessayer."

client/ui/i18n/locales/hu/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Keresse a NetBirdöt a tálcán"
     },
+    "welcome.titleMac": {
+        "message": "Keresse a NetBirdöt a menüsorban"
+    },
     "welcome.description": {
         "message": "A NetBird a tálcán fut. Kattintson az ikonra a csatlakozáshoz, profilváltáshoz vagy a beállítások megnyitásához."
     },
+    "welcome.descriptionMac": {
+        "message": "A NetBird a menüsorban fut. Kattintson az ikonra a csatlakozáshoz, profilváltáshoz vagy a beállítások megnyitásához."
+    },
     "welcome.continue": {
         "message": "Folytatás"
     },
@@ -1293,10 +1299,13 @@
         "message": "Dokumentáció"
     },
     "daemon.outdated.title": {
-        "message": "A NetBird szolgáltatás elavult"
+        "message": "A NetBird Kliens elavult"
     },
     "daemon.outdated.description": {
-        "message": "Frissítsd a NetBird szolgáltatást az alkalmazás használatához."
+        "message": "Az új GUI nem kompatibilis a régebbi klienseddel. Frissítsd a klienst az új alkalmazás használatához."
+    },
+    "daemon.outdated.download": {
+        "message": "Legújabb letöltése"
     },
     "error.jwt_clock_skew": {
         "message": "A bejelentkezés sikertelen: az eszköz órája eltér a szerverétől. Kérjük, szinkronizálja a rendszer óráját, majd próbálja újra."

client/ui/i18n/locales/it/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Cerchi NetBird nella tray"
     },
+    "welcome.titleMac": {
+        "message": "Cerchi NetBird nella barra dei menu"
+    },
     "welcome.description": {
         "message": "NetBird risiede nella tray. Clicchi sull'icona per connettersi, cambiare profilo o aprire le impostazioni."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird risiede nella barra dei menu. Clicchi sull'icona per connettersi, cambiare profilo o aprire le impostazioni."
+    },
     "welcome.continue": {
         "message": "Continua"
     },
@@ -1293,10 +1299,13 @@
         "message": "Documentazione"
     },
     "daemon.outdated.title": {
-        "message": "Il servizio NetBird è obsoleto"
+        "message": "NetBird Client è obsoleto"
     },
     "daemon.outdated.description": {
-        "message": "Aggiorna il servizio NetBird per usare questa app."
+        "message": "La nuova GUI non è compatibile con il tuo client precedente. Aggiorna il client per usare la nuova applicazione."
+    },
+    "daemon.outdated.download": {
+        "message": "Scarica l'ultima versione"
     },
     "error.jwt_clock_skew": {
         "message": "Accesso non riuscito: l'orologio di questo dispositivo non è sincronizzato con il server. Sincronizzi l'orologio di sistema e riprovi."

client/ui/i18n/locales/pt/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Procure o NetBird na sua bandeja"
     },
+    "welcome.titleMac": {
+        "message": "Procure o NetBird na sua barra de menus"
+    },
     "welcome.description": {
         "message": "O NetBird fica na sua bandeja. Clique no ícone para conectar, alternar perfis ou abrir as configurações."
     },
+    "welcome.descriptionMac": {
+        "message": "O NetBird fica na sua barra de menus. Clique no ícone para conectar, alternar perfis ou abrir as configurações."
+    },
     "welcome.continue": {
         "message": "Continuar"
     },
@@ -1293,10 +1299,13 @@
         "message": "Documentação"
     },
     "daemon.outdated.title": {
-        "message": "O serviço NetBird está desatualizado"
+        "message": "O NetBird Client está desatualizado"
     },
     "daemon.outdated.description": {
-        "message": "Atualize o serviço NetBird para usar este aplicativo."
+        "message": "A nova GUI não é compatível com o seu cliente mais antigo. Atualize o seu cliente para usar o novo aplicativo."
+    },
+    "daemon.outdated.download": {
+        "message": "Baixar a versão mais recente"
     },
     "error.jwt_clock_skew": {
         "message": "Falha no login: o relógio deste dispositivo está fora de sincronia com o servidor. Sincronize o relógio do sistema e tente novamente."

client/ui/i18n/locales/ru/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "Найдите NetBird в системном трее"
     },
+    "welcome.titleMac": {
+        "message": "Найдите NetBird в строке меню"
+    },
     "welcome.description": {
         "message": "NetBird находится в системном трее. Нажмите на значок, чтобы подключиться, переключить профиль или открыть настройки."
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird находится в строке меню. Нажмите на значок, чтобы подключиться, переключить профиль или открыть настройки."
+    },
     "welcome.continue": {
         "message": "Продолжить"
     },
@@ -1293,10 +1299,13 @@
         "message": "Документация"
     },
     "daemon.outdated.title": {
-        "message": "Служба NetBird устарела"
+        "message": "Клиент NetBird устарел"
     },
     "daemon.outdated.description": {
-        "message": "Обновите службу NetBird, чтобы использовать это приложение."
+        "message": "Новый GUI несовместим с вашим более старым клиентом. Обновите клиент, чтобы использовать новое приложение."
+    },
+    "daemon.outdated.download": {
+        "message": "Скачать последнюю версию"
     },
     "error.jwt_clock_skew": {
         "message": "Не удалось войти: часы этого устройства рассинхронизированы с сервером. Синхронизируйте системные часы и повторите попытку."

client/ui/i18n/locales/zh-CN/common.json

@@ -1034,9 +1034,15 @@
     "welcome.title": {
         "message": "在托盘中查找 NetBird"
     },
+    "welcome.titleMac": {
+        "message": "在菜单栏中查找 NetBird"
+    },
     "welcome.description": {
         "message": "NetBird 驻留在您的托盘中。点击图标即可连接、切换配置文件或打开设置。"
     },
+    "welcome.descriptionMac": {
+        "message": "NetBird 驻留在您的菜单栏中。点击图标即可连接、切换配置文件或打开设置。"
+    },
     "welcome.continue": {
         "message": "继续"
     },
@@ -1293,10 +1299,13 @@
         "message": "文档"
     },
     "daemon.outdated.title": {
-        "message": "NetBird 服务版本过旧"
+        "message": "NetBird 客户端版本过旧"
     },
     "daemon.outdated.description": {
-        "message": "请更新 NetBird 服务以使用此应用。"
+        "message": "新版 GUI 与您较旧的客户端不兼容。请更新客户端以使用新应用。"
+    },
+    "daemon.outdated.download": {
+        "message": "下载最新版本"
     },
     "error.jwt_clock_skew": {
         "message": "登录失败：此设备的时钟与服务器不同步。请同步您的系统时钟后重试。"

client/ui/services/profileswitcher.go

@@ -12,13 +12,15 @@ import (
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 )
 
-// ProfileSwitcher holds the reconnect policy shared by the tray and React
-// frontend so both flip profiles identically. The policy keys off prevStatus
-// from DaemonFeed.Get at SwitchActive entry:
+// ProfileSwitcher holds the switch policy shared by the tray and React
+// frontend so both flip profiles identically. SwitchActive (plain selection:
+// header dropdown, tray submenu) always connects after the switch;
+// SwitchActiveNoConnect (manage-profiles screen) never does, so the user can
+// still adjust the management URL before connecting. prevStatus from
+// DaemonFeed.Get at entry only decides the teardown:
 //
-//	Connected/Connecting → Switch + Down + Up; optimistic Connecting paint.
-//	NeedsLogin/LoginFailed/SessionExpired → Switch + Down; clear stale error for re-login.
-//	Idle → Switch only.
+//	Connected/Connecting/NeedsLogin/LoginFailed/SessionExpired → Down first.
+//	Idle → no Down.
 type ProfileSwitcher struct {
 	profiles   *Profiles
 	connection *Connection
@@ -29,29 +31,40 @@ func NewProfileSwitcher(profiles *Profiles, connection *Connection, feed *Daemon
 	return &ProfileSwitcher{profiles: profiles, connection: connection, feed: feed}
 }
 
-// SwitchActive switches to the named profile applying the reconnect policy.
+// SwitchActive switches to the named profile and always connects afterwards.
 func (s *ProfileSwitcher) SwitchActive(ctx context.Context, p ProfileRef) error {
+	return s.switchActive(ctx, p, true)
+}
+
+// SwitchActiveNoConnect switches to the named profile without connecting,
+// tearing down any existing connection first.
+func (s *ProfileSwitcher) SwitchActiveNoConnect(ctx context.Context, p ProfileRef) error {
+	return s.switchActive(ctx, p, false)
+}
+
+func (s *ProfileSwitcher) switchActive(ctx context.Context, p ProfileRef, connect bool) error {
 	prevStatus := ""
-	if st, err := s.feed.Get(ctx); err == nil {
-		prevStatus = st.Status
-	} else {
-		log.Warnf("profileswitcher: get status: %v", err)
+	if s.feed != nil {
+		if st, err := s.feed.Get(ctx); err == nil {
+			prevStatus = st.Status
+		} else {
+			log.Warnf("profileswitcher: get status: %v", err)
+		}
 	}
 
-	wasActive := strings.EqualFold(prevStatus, StatusConnected) ||
-		strings.EqualFold(prevStatus, StatusConnecting)
-	needsDown := wasActive ||
+	needsDown := strings.EqualFold(prevStatus, StatusConnected) ||
+		strings.EqualFold(prevStatus, StatusConnecting) ||
 		strings.EqualFold(prevStatus, StatusNeedsLogin) ||
 		strings.EqualFold(prevStatus, StatusLoginFailed) ||
 		strings.EqualFold(prevStatus, StatusSessionExpired)
 
-	log.Infof("profileswitcher: switch profile=%q prevStatus=%q wasActive=%v needsDown=%v",
-		p.ProfileName, prevStatus, wasActive, needsDown)
+	log.Infof("profileswitcher: switch profile=%q prevStatus=%q connect=%v needsDown=%v",
+		p.ProfileName, prevStatus, connect, needsDown)
 
-	// Optimistic Connecting paint only when wasActive: those prevStatuses emit
-	// stale Connected + transient Idle pushes during Down that must be
-	// suppressed until Up resumes the stream (see DaemonFeed suppression table).
-	if wasActive {
+	// Optimistic Connecting paint plus stale-push suppression during Down (see
+	// DaemonFeed suppression table); also arms the login-watch that pops
+	// browser-login when the new profile turns out to need SSO.
+	if connect && s.feed != nil {
 		s.feed.BeginProfileSwitch()
 	}
 
@@ -76,9 +89,9 @@ func (s *ProfileSwitcher) SwitchActive(ctx context.Context, p ProfileRef) error
 		}
 	}
 
-	if wasActive {
+	if connect {
 		if err := s.connection.Up(ctx, UpParams(p)); err != nil {
-			return fmt.Errorf("reconnect %q: %w", p.ProfileName, err)
+			return fmt.Errorf("connect %q: %w", p.ProfileName, err)
 		}
 	}
 

client/ui/services/windowmanager.go

@@ -185,37 +185,38 @@ func (s *WindowManager) OpenBrowserLogin(uri string) {
 			startURL = "/#/dialog/browser-login?uri=" + url.QueryEscape(uri)
 		}
 		s.hideOtherWindowsLocked("browser-login")
-		// Prefer the main window's screen (multi-monitor); falls back to OS-default centering.
-		var screen *application.Screen
-		if s.mainWindow != nil {
-			if sc, err := s.mainWindow.GetScreen(); err == nil {
-				screen = sc
-			}
-		}
 		opts := DialogWindowOptions("browser-login", s.title("window.title.signIn"), startURL, s.linuxIcon)
 		// Not always-on-top: it would obscure the browser tab the user logs in through.
 		opts.AlwaysOnTop = false
 		opts.InitialPosition = application.WindowCentered
-		opts.Screen = screen
+		// Open on the active (where users cursor is) display, like the session-expiration dialog.
+		opts.Screen = s.getScreenBasedOnCursorPosition()
 		s.browserLogin = s.app.Window.NewWithOptions(opts)
 		bl := s.browserLogin
-		// Red-X close means cancel: emit the event so startLogin() tears down the SSO wait.
 		bl.OnWindowEvent(events.Common.WindowClosing, func(_ *application.WindowEvent) {
-			s.app.Event.Emit(EventBrowserLoginCancel)
 			s.mu.Lock()
-			s.browserLogin = nil
-			s.restoreHiddenWindowsLocked()
+			// Only a live user red-X still has this registered; programmatic closers
+			// nil s.browserLogin first and clean up themselves. Guarding here stops a
+			// stale close event from wiping a replacement popup's state.
+			userClosed := s.browserLogin == bl
+			if userClosed {
+				s.browserLogin = nil
+				s.restoreHiddenWindowsLocked()
+			}
 			s.mu.Unlock()
+			if userClosed {
+				s.app.Event.Emit(EventBrowserLoginCancel)
+			}
 		})
-		s.centerWhenReady(s.browserLogin)
+		s.centerOnCursorScreen(s.browserLogin)
 		return
 	}
 	if uri != "" {
 		s.browserLogin.SetURL("/#/dialog/browser-login?uri=" + url.QueryEscape(uri))
 	}
+	s.centerOnCursorScreen(s.browserLogin)
 	s.browserLogin.Show()
 	s.browserLogin.Focus()
-	s.centerWhenReady(s.browserLogin)
 }
 
 // BrowserLoginWindow returns the live SSO popup, or nil. While non-nil it is the
@@ -238,6 +239,15 @@ func (s *WindowManager) CloseBrowserLogin() {
 	s.mu.Lock()
 	w := s.browserLogin
 	s.browserLogin = nil
+	// The WindowClosing hook no-ops on a programmatic close, so restore here —
+	// but only if a popup was actually open. The frontend calls this even when no
+	// popup was ever shown (e.g. resetDialog() after an early RequestExtend failure,
+	// or connection.ts's catch path), and hiddenForLogin is shared with
+	// OpenInstallProgress, so an unconditional restore could re-show windows a
+	// still-running install-progress is hiding.
+	if w != nil {
+		s.restoreHiddenWindowsLocked()
+	}
 	s.mu.Unlock()
 	if w != nil {
 		w.Close()
@@ -279,6 +289,35 @@ func (s *WindowManager) CloseSessionExpiration() {
 	}
 }
 
+// CloseRenewFlow tears down the SSO session-renewal UI in a single call: it
+// closes the browser-login popup and the session-expiration window together.
+func (s *WindowManager) CloseRenewFlow() {
+	s.mu.Lock()
+	bl := s.browserLogin
+	se := s.sessionExpiration
+	s.browserLogin = nil
+	s.sessionExpiration = nil
+	if se != nil {
+		kept := s.hiddenForLogin[:0]
+		for _, w := range s.hiddenForLogin {
+			if w != se {
+				kept = append(kept, w)
+			}
+		}
+		s.hiddenForLogin = kept
+	}
+	s.restoreHiddenWindowsLocked()
+	s.mu.Unlock()
+
+	// Close after unlock so the re-entrant handlers can take s.mu.
+	if bl != nil {
+		bl.Close()
+	}
+	if se != nil {
+		se.Close()
+	}
+}
+
 // OpenInstallProgress shows the install-progress window and hides the rest for the duration
 // (restored on close). It owns its own result polling since the daemon restarts mid-install.
 func (s *WindowManager) OpenInstallProgress(version string) {

client/ui/tray.go

@@ -30,6 +30,8 @@ const (
 
 	statusError = "Error"
 
+	quitDownTimeout = 5 * time.Second
+
 	urlGitHubRepo     = "https://github.com/netbirdio/netbird"
 	urlGitHubReleases = "https://github.com/netbirdio/netbird/releases/latest"
 	urlDocs           = "https://docs.netbird.io"
@@ -446,11 +448,28 @@ func (t *Tray) buildMenu() *application.Menu {
 	menu.AddSeparator()
 	menu.Add(t.loc.T("tray.menu.quit")).
 		SetAccelerator("CmdOrCtrl+Q").
-		OnClick(func(*application.Context) { t.app.Quit() })
+		OnClick(func(*application.Context) { t.handleQuit() })
 
 	return menu
 }
 
+func (t *Tray) handleQuit() {
+	t.profileMu.Lock()
+	if t.switchCancel != nil {
+		t.switchCancel()
+		t.switchCancel = nil
+	}
+	t.profileMu.Unlock()
+	t.svc.DaemonFeed.CancelProfileSwitch()
+
+	ctx, cancel := context.WithTimeout(context.Background(), quitDownTimeout)
+	defer cancel()
+	if err := t.svc.Connection.Down(ctx); err != nil {
+		log.Errorf("disconnect on quit: %v", err)
+	}
+	t.app.Quit()
+}
+
 // handleConnect receives the clicked item from the buildMenu closure —
 // t.upItem is menuMu-guarded and must not be read here.
 func (t *Tray) handleConnect(upItem *application.MenuItem) {

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

@@ -91,14 +91,22 @@ func availableProviders() []providerCase {
 		if region == "" {
 			region = "us-east-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.WireBedrock})
+		// A valid Bedrock inference-profile id (region prefix + date + version),
+		// overridable per account. `global.` profiles can be invoked from any
+		// region; set AWS_BEDROCK_MODEL to match the enabled profile for the token.
+		model := os.Getenv("AWS_BEDROCK_MODEL")
+		if model == "" {
+			model = "global.anthropic.claude-haiku-4-5-20251001-v1:0"
+		}
+		ps = append(ps, providerCase{name: "bedrock", catalogID: "bedrock_api", upstream: "https://bedrock-runtime." + region + ".amazonaws.com", apiKey: k, model: model, kind: harness.WireBedrock})
 	}
 	return ps
 }
 
 // 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).
+// a uniquely-priced model registered under the id an operator would paste —
+// the normalized catalog id for Bedrock, the raw form elsewhere (including the
+// "@version" Vertex id, which the router must normalize to route).
 func providerRequest(pc providerCase) api.AgentNetworkProviderRequest {
 	req := api.AgentNetworkProviderRequest{
 		Name:        pc.name,
@@ -107,10 +115,19 @@ func providerRequest(pc providerCase) api.AgentNetworkProviderRequest {
 		ApiKey:      &pc.apiKey,
 		Enabled:     ptr(true),
 	}
-	if pc.kind != harness.WireVertex {
-		req.Models = &[]api.AgentNetworkProviderModel{
-			{Id: pc.model, InputPer1k: 0.001, OutputPer1k: 0.002},
-		}
+	// The router matches the normalized catalog id. Bedrock's request model
+	// travels as a region-prefixed inference-profile id in the URL path
+	// (us.anthropic...), which the router strips before matching, so register
+	// the normalized form here or routing fails as model_not_routable. Vertex is
+	// registered with the raw "@version" id the Google console documents — the
+	// router normalizes the registered id, so the operator can paste either
+	// form; keeping the raw form here guards that normalization end to end.
+	modelID := pc.model
+	if pc.kind == harness.WireBedrock {
+		modelID = catalogModel(pc)
+	}
+	req.Models = &[]api.AgentNetworkProviderModel{
+		{Id: modelID, InputPer1k: 0.001, OutputPer1k: 0.002},
 	}
 	return req
 }
@@ -201,11 +218,13 @@ func TestProvidersMatrix(t *testing.T) {
 	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
 
 	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Resolve first: the DNS lookup triggers the lazy-connection warm-up, waking
+	// the proxy peer so WaitProxyPeer then observes it connected.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve agent-network endpoint to proxy IP")
 	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 matrix {
 		pc := pc

e2e/agentnetwork/guardrail_test.go

@@ -4,6 +4,7 @@ package agentnetwork
 
 import (
 	"context"
+	"regexp"
 	"strings"
 	"testing"
 	"time"
@@ -15,13 +16,29 @@ import (
 	"github.com/netbirdio/netbird/shared/management/http/api"
 )
 
+// bedrockRegionPrefixes and bedrockVersionSuffix mirror the proxy's Bedrock
+// model normalization (region/inference-profile prefix + version suffix) so the
+// provider is registered under the same catalog key the router matches against.
+var (
+	bedrockRegionPrefixes = []string{"us.", "eu.", "apac.", "global."}
+	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).
+// path-routed provider's configured model — the form the router and guardrail
+// allowlist compare against (Bedrock region prefix + version stripped, Vertex
+// @version stripped).
 func catalogModel(pc providerCase) string {
 	switch pc.kind {
 	case harness.WireBedrock:
-		return strings.TrimPrefix(pc.model, "us.")
+		m := pc.model
+		for _, p := range bedrockRegionPrefixes {
+			if strings.HasPrefix(m, p) {
+				m = m[len(p):]
+				break
+			}
+		}
+		return bedrockVersionSuffix.ReplaceAllString(m, "")
 	case harness.WireVertex:
 		return strings.SplitN(pc.model, "@", 2)[0]
 	default:
@@ -108,7 +125,15 @@ func TestModelAllowlistEnforced(t *testing.T) {
 		require.NoError(t, perr, "create provider %s", pc.name)
 		id := prov.Id
 		ids = append(ids, id)
-		allowed = append(allowed, catalogModel(pc))
+		// Vertex allowlists the raw "@version" id an operator copies from the
+		// Google console; the guardrail must normalize the entry to match the
+		// version-stripped request model. Other providers allowlist the
+		// normalized catalog id.
+		if pc.kind == harness.WireVertex {
+			allowed = append(allowed, pc.model)
+		} else {
+			allowed = append(allowed, catalogModel(pc))
+		}
 		t.Cleanup(func() { _ = srv.DeleteProvider(context.Background(), id) })
 	}
 
@@ -147,11 +172,13 @@ func TestModelAllowlistEnforced(t *testing.T) {
 	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
 
 	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Resolve first: the DNS lookup triggers the lazy-connection warm-up, waking
+	// the proxy peer so WaitProxyPeer then observes it connected.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve agent-network endpoint to proxy IP")
 	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
@@ -163,6 +190,25 @@ func TestModelAllowlistEnforced(t *testing.T) {
 			// 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)
+
+			if pc.kind != harness.WireVertex {
+				return
+			}
+			// The unversioned model id — the shape an Anthropic SDK client sends
+			// when configured without an "@version" suffix (customer-reported) —
+			// must pass the same allowlist entry and resolve upstream.
+			assert.Equal(t, 200, sendModel(ctx, t, cl, settings.Endpoint, proxyIP, pc, catalogModel(pc)),
+				"unversioned model id must be permitted for %s", pc.name)
+			// Token counting: the URL carries the literal "count-tokens"
+			// pseudo-model and the real model travels in the body (Claude Code
+			// sends one such call per request). Allowed body model → served.
+			code, _, err := cl.VertexCountTokens(ctx, settings.Endpoint, proxyIP, pc.project, pc.region, pc.model)
+			require.NoError(t, err, "count-tokens request must reach the proxy for %s", pc.name)
+			assert.Equal(t, 200, code, "count-tokens with an allowlisted body model must be permitted for %s", pc.name)
+			// Disallowed body model → denied before the upstream.
+			code, _, err = cl.VertexCountTokens(ctx, settings.Endpoint, proxyIP, pc.project, pc.region, disallowedModel(pc))
+			require.NoError(t, err, "count-tokens request must reach the proxy for %s", pc.name)
+			assert.Equal(t, 403, code, "count-tokens with a body model outside the allowlist must be denied for %s", pc.name)
 		})
 	}
 }

e2e/agentnetwork/skiptls_test.go

@@ -104,11 +104,13 @@ func TestProviderSkipTLSVerification(t *testing.T) {
 	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
 
 	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Resolve first: the DNS lookup triggers the lazy-connection warm-up, waking
+	// the proxy peer so WaitProxyPeer then observes it connected.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
 	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 endpoint to proxy IP")
 
 	// Positive: skip=true reaches the self-signed upstream. Retry to absorb
 	// tunnel/DNS jitter on the first call; success also proves the path works.

e2e/agentnetwork/vllm_test.go

@@ -106,11 +106,13 @@ func TestVLLMProvider(t *testing.T) {
 	t.Cleanup(func() { _ = cl.Terminate(context.Background()) })
 
 	require.NoError(t, cl.WaitConnected(ctx, 90*time.Second), "client must connect to management")
+	// Resolve first: the DNS lookup triggers the lazy-connection warm-up, waking
+	// the proxy peer so WaitProxyPeer then observes it connected.
+	proxyIP, err := cl.ResolveProxyIP(ctx, settings.Endpoint)
+	require.NoError(t, err, "resolve endpoint to proxy IP")
 	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 endpoint to proxy IP")
 
 	before, _ := srv.ListAccessLogs(ctx)
 	sessionID := "e2e-session-vllm"

e2e/harness/client.go

@@ -231,6 +231,17 @@ func (cl *Client) Vertex(ctx context.Context, endpoint, proxyIP, project, region
 	return cl.post(ctx, endpoint, proxyIP, path, body, withSessionID(nil, sessionID))
 }
 
+// VertexCountTokens issues the Anthropic-on-Vertex token-count POST over the
+// tunnel. The URL carries the literal "count-tokens" pseudo-model and the real
+// model travels in the body — the one Vertex shape whose model is body-routed,
+// sent by Claude Code before each request. The proxy must resolve the body
+// model for the guardrail allowlist and routing.
+func (cl *Client) VertexCountTokens(ctx context.Context, endpoint, proxyIP, project, region, model string) (int, string, error) {
+	path := fmt.Sprintf("/v1/projects/%s/locations/%s/publishers/anthropic/models/count-tokens:rawPredict", project, region)
+	body := fmt.Sprintf(`{"model":%q,"messages":[{"role":"user","content":"hi"}]}`, model)
+	return cl.post(ctx, endpoint, proxyIP, path, body, nil)
+}
+
 // 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

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

go.sum

@@ -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=

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",

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")
+}

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/vertex_model.go

@@ -0,0 +1,17 @@
+package llm
+
+import "strings"
+
+// NormalizeVertexModel strips the "@version" suffix from a Google Vertex AI
+// publisher model id so it matches the catalog/pricing key, e.g.
+// "claude-sonnet-4-5@20250929" -> "claude-sonnet-4-5". It is the single source
+// of truth shared by the request parser (which normalizes the request model
+// from the URL path), the router (which normalizes the operator's registered
+// Vertex model ids), and the guardrail (which normalizes allowlist entries) so
+// both spellings of the same model compare equal.
+func NormalizeVertexModel(modelID string) string {
+	if at := strings.Index(modelID, "@"); at >= 0 {
+		return modelID[:at]
+	}
+	return modelID
+}

proxy/internal/llm/vertex_model_test.go

@@ -0,0 +1,23 @@
+package llm
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+func TestNormalizeVertexModel(t *testing.T) {
+	cases := map[string]string{
+		"claude-sonnet-4-5@20250929": "claude-sonnet-4-5",
+		"claude-opus-4-6@20250514":   "claude-opus-4-6",
+		// Bare ids (the form Vertex resolves to the default version) pass through.
+		"claude-opus-4-6": "claude-opus-4-6",
+		// Numeric version aliases used by Google publisher models.
+		"text-embedding-005@001": "text-embedding-005",
+		// A leading "@" yields the empty string; callers treat that as "no model".
+		"@cf/meta/llama-3-8b": "",
+	}
+	for in, want := range cases {
+		require.Equal(t, want, NormalizeVertexModel(in), "normalize %q", in)
+	}
+}

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

@@ -13,6 +13,7 @@ import (
 	"context"
 	"unicode/utf8"
 
+	"github.com/netbirdio/netbird/proxy/internal/llm"
 	"github.com/netbirdio/netbird/proxy/internal/middleware"
 )
 
@@ -162,6 +163,15 @@ func (m *Middleware) modelInAllowlist(model string) bool {
 		if allowed == normalised {
 			return true
 		}
+		// Vertex model ids are often stored with their "@version" suffix (e.g.
+		// "claude-opus-4-6@20250514") — the id the Google console documents —
+		// while the request parser strips the suffix before the model reaches
+		// the guardrail. Accept an entry whose version-stripped form matches so
+		// both spellings of the same model pass one allowlist. Entries without
+		// an "@" are untouched, so non-Vertex matching stays exact.
+		if v := llm.NormalizeVertexModel(allowed); v != "" && v != allowed && v == normalised {
+			return true
+		}
 	}
 	return false
 }

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

@@ -70,6 +70,27 @@ func TestAllowlistMatchAllows(t *testing.T) {
 	assert.Equal(t, middleware.DecisionAllow, out.Decision, "model in allowlist must be allowed")
 }
 
+// A Vertex allowlist entry stored with its "@version" suffix must match the
+// request model, which reaches the guardrail with the suffix already stripped
+// by the request parser.
+func TestAllowlistVertexVersionedEntryMatchesStrippedModel(t *testing.T) {
+	mw := New(Config{ModelAllowlist: []string{"claude-opus-4-6@20250514"}})
+	out, err := mw.Invoke(context.Background(), newInput(
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4-6"},
+	))
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision,
+		"@version allowlist entry must match the version-stripped request model")
+
+	// The stripped form must not over-match a different model.
+	out, err = mw.Invoke(context.Background(), newInput(
+		middleware.KV{Key: middleware.KeyLLMModel, Value: "claude-opus-4-8"},
+	))
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision,
+		"a different model must stay denied")
+}
+
 func TestAllowlistMissDenies(t *testing.T) {
 	mw := New(Config{ModelAllowlist: []string{"gpt-4o"}})
 	out, err := mw.Invoke(context.Background(), newInput(

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

@@ -104,3 +104,96 @@ func TestModelAllowlist_URLRoutedProviders(t *testing.T) {
 		})
 	}
 }
+
+// TestModelAllowlist_VertexRequestShapes replays the Vertex request shapes an
+// Anthropic SDK client (Claude Code) actually sends — the model travels in the
+// URL path, optionally without an "@version" suffix, at the global location —
+// against allowlists holding either the bare catalog id or the raw versioned
+// id. The URLs mirror the shape of a customer-reported request (unversioned
+// model, global location, a project id that itself contains "claude").
+func TestModelAllowlist_VertexRequestShapes(t *testing.T) {
+	const (
+		opusBare      = "/v1/projects/corp-gcp-it-all-claude/locations/global/publishers/anthropic/models/claude-opus-4-6:rawPredict"
+		opusBareSSE   = "/v1/projects/corp-gcp-it-all-claude/locations/global/publishers/anthropic/models/claude-opus-4-6:streamRawPredict"
+		countTokens   = "/v1/projects/corp-gcp-it-all-claude/locations/global/publishers/anthropic/models/count-tokens:rawPredict"
+		messagesBody  = `{"anthropic_version":"vertex-2023-10-16","messages":[{"role":"user","content":"hi"}]}`
+		countOpusBody = `{"model":"claude-opus-4-6","messages":[{"role":"user","content":"hi"}]}`
+	)
+
+	tests := []struct {
+		name      string
+		url       string
+		body      string
+		allowlist []string
+		decision  middleware.Decision
+		denyCode  string
+	}{
+		{
+			name:      "unversioned model allowed by bare catalog entry",
+			url:       opusBare,
+			body:      messagesBody,
+			allowlist: []string{"claude-opus-4-6"},
+			decision:  middleware.DecisionAllow,
+		},
+		{
+			name:      "unversioned model allowed by @version allowlist entry",
+			url:       opusBare,
+			body:      messagesBody,
+			allowlist: []string{"claude-opus-4-6@20250514"},
+			decision:  middleware.DecisionAllow,
+		},
+		{
+			name:      "streaming action allowed the same as rawPredict",
+			url:       opusBareSSE,
+			body:      messagesBody,
+			allowlist: []string{"claude-opus-4-6"},
+			decision:  middleware.DecisionAllow,
+		},
+		{
+			// The original customer report: a Sonnet-only allowlist must block
+			// an Opus request regardless of the id carrying a version suffix.
+			name:      "unversioned model outside the allowlist denied",
+			url:       opusBare,
+			body:      messagesBody,
+			allowlist: []string{"claude-sonnet-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+		{
+			name:      "count-tokens resolves the body model and passes when allowed",
+			url:       countTokens,
+			body:      countOpusBody,
+			allowlist: []string{"claude-opus-4-6"},
+			decision:  middleware.DecisionAllow,
+		},
+		{
+			name:      "count-tokens with a disallowed body model denied",
+			url:       countTokens,
+			body:      countOpusBody,
+			allowlist: []string{"claude-sonnet-4-5"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+		{
+			// No model in the body: the pseudo-model stays and fails closed.
+			name:      "count-tokens without a body model fails closed",
+			url:       countTokens,
+			body:      messagesBody,
+			allowlist: []string{"claude-opus-4-6"},
+			decision:  middleware.DecisionDeny,
+			denyCode:  "llm_policy.model_blocked",
+		},
+	}
+
+	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_request_parser/middleware.go

@@ -253,9 +253,7 @@ func parseVertexPath(reqPath string) (vertexRequest, bool) {
 	if c := strings.LastIndex(rest, ":"); c >= 0 {
 		model, action = rest[:c], rest[c+1:]
 	}
-	if at := strings.Index(model, "@"); at >= 0 {
-		model = model[:at]
-	}
+	model = llm.NormalizeVertexModel(model)
 	if model == "" {
 		return vertexRequest{}, false
 	}
@@ -276,24 +274,44 @@ func vertexPublisherVendor(publisher string) string {
 	}
 }
 
+// vertexCountTokensModel is the pseudo-model id of the Vertex token-count
+// endpoint (…/models/count-tokens:rawPredict). It is the one Vertex request
+// shape whose real model travels in the JSON body rather than the URL path.
+const vertexCountTokensModel = "count-tokens"
+
 // invokeVertex emits the model/vendor/session/prompt for a Vertex publisher
 // request, using the publisher's parser to read the (vendor-native) body.
 func (m middlewareImpl) invokeVertex(in *middleware.Input, vx vertexRequest) *middleware.Output {
 	out := &middleware.Output{Decision: middleware.DecisionAllow}
 	vendor := vertexPublisherVendor(vx.publisher)
 
-	md := []middleware.KV{}
-	if vendor != "" {
-		md = append(md, middleware.KV{Key: middleware.KeyLLMProvider, Value: vendor})
-	}
-	md = append(md, middleware.KV{Key: middleware.KeyLLMModel, Value: vx.model})
-	md = append(md, middleware.KV{Key: middleware.KeyLLMStream, Value: strconv.FormatBool(vx.stream)})
-
 	var parser llm.Parser
 	if vendor != "" {
 		parser, _ = llm.ParserByName(vendor)
 	}
 
+	model := vx.model
+	// Token counting carries the literal "count-tokens" pseudo-model in the URL
+	// and the real model in the body (Claude Code issues one such call per
+	// request). Resolve the body model so the guardrail allowlist and the
+	// router evaluate the actual model — otherwise every count-tokens call
+	// under an allowlist denies as model_blocked. A body the parser cannot
+	// read keeps the pseudo-model, which fails closed downstream.
+	if model == vertexCountTokensModel && parser != nil {
+		if facts, err := parser.ParseRequest(in.Body); err == nil && facts.Model != "" {
+			if bodyModel := llm.NormalizeVertexModel(facts.Model); bodyModel != "" {
+				model = bodyModel
+			}
+		}
+	}
+
+	md := []middleware.KV{}
+	if vendor != "" {
+		md = append(md, middleware.KV{Key: middleware.KeyLLMProvider, Value: vendor})
+	}
+	md = append(md, middleware.KV{Key: middleware.KeyLLMModel, Value: model})
+	md = append(md, middleware.KV{Key: middleware.KeyLLMStream, Value: strconv.FormatBool(vx.stream)})
+
 	sessionID := sessionIDFromHeaders(in.Headers)
 	if sessionID == "" && parser != nil {
 		sessionID = parser.ExtractSessionID(in.Body)

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 +556,22 @@ 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
+		}
+		// Vertex request models likewise reach the router with the "@version"
+		// suffix already stripped by the parser, while the operator may register
+		// the raw versioned id the Google console documents (e.g.
+		// "claude-opus-4-6@20250514"). Normalize the candidate so both spellings
+		// of the same model match.
+		if route.Vertex && llm.NormalizeVertexModel(candidate) == model {
+			return true
+		}
 	}
 	return false
 }

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

@@ -0,0 +1,64 @@
+package llm_router
+
+import (
+	"context"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
+)
+
+// TestRouteClaimsModel_VertexNormalizesCandidate guards the Vertex analog of
+// the native-Bedrock routing gap: the request model reaches the router already
+// normalized (the parser strips the "@version" suffix from the URL path), so a
+// provider registered with the raw versioned id must still match.
+func TestRouteClaimsModel_VertexNormalizesCandidate(t *testing.T) {
+	route := ProviderRoute{Vertex: true, Models: []string{"claude-opus-4-6@20250514"}}
+	assert.True(t, routeClaimsModel(route, "claude-opus-4-6"),
+		"raw @version Vertex model must match the normalized request model")
+	assert.False(t, routeClaimsModel(route, "claude-haiku-4-5"),
+		"a model outside the provider's list must not match")
+
+	bare := ProviderRoute{Vertex: true, Models: []string{"claude-opus-4-6"}}
+	assert.True(t, routeClaimsModel(bare, "claude-opus-4-6"),
+		"a bare catalog id keeps matching exactly")
+
+	// Non-Vertex routes keep exact matching for "@"-suffixed candidates.
+	direct := ProviderRoute{Models: []string{"claude-opus-4-6@20250514"}}
+	assert.False(t, routeClaimsModel(direct, "claude-opus-4-6"),
+		"non-Vertex routes must not normalize @version candidates")
+}
+
+// TestRouter_VertexUnversionedModelRoutes replays the customer-reported request
+// end to end through the router: an unversioned model id at the global location
+// (…/models/claude-opus-4-6:rawPredict) must route on a Vertex provider whose
+// models were registered with the raw "@version" ids.
+func TestRouter_VertexUnversionedModelRoutes(t *testing.T) {
+	route := vertexRoute()
+	route.Models = []string{"claude-opus-4-6@20250514", "claude-sonnet-4-5@20250929"}
+	mw := New(Config{Providers: []ProviderRoute{route}})
+
+	in := pathRoutedInput(
+		"/v1/projects/corp-gcp-it-all-claude/locations/global/publishers/anthropic/models/claude-opus-4-6:rawPredict",
+		"anthropic",
+		"claude-opus-4-6",
+	)
+	out, err := mw.Invoke(context.Background(), in)
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionAllow, out.Decision,
+		"unversioned request model must route on a provider registered with @version ids")
+
+	denied := pathRoutedInput(
+		"/v1/projects/corp-gcp-it-all-claude/locations/global/publishers/anthropic/models/claude-haiku-4-5:rawPredict",
+		"anthropic",
+		"claude-haiku-4-5",
+	)
+	out, err = mw.Invoke(context.Background(), denied)
+	require.NoError(t, err)
+	assert.Equal(t, middleware.DecisionDeny, out.Decision,
+		"a model outside the provider's registered list must still deny")
+	require.NotNil(t, out.DenyReason)
+	assert.Equal(t, denyCodeNotRoutable, out.DenyReason.Code)
+}

proxy/internal/proxy/agent_network_vertex_guardrail_realstack_test.go

@@ -0,0 +1,291 @@
+package proxy_test
+
+import (
+	"context"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
+	"runtime"
+	"strings"
+	"sync/atomic"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"google.golang.org/grpc"
+	"google.golang.org/grpc/credentials/insecure"
+	"google.golang.org/grpc/test/bufconn"
+
+	"github.com/netbirdio/netbird/management/internals/modules/agentnetwork"
+	agentNetworkTypes "github.com/netbirdio/netbird/management/internals/modules/agentnetwork/types"
+	rpservice "github.com/netbirdio/netbird/management/internals/modules/reverseproxy/service"
+	mgmtgrpc "github.com/netbirdio/netbird/management/internals/shared/grpc"
+	"github.com/netbirdio/netbird/management/server/store"
+	"github.com/netbirdio/netbird/proxy/internal/middleware"
+	"github.com/netbirdio/netbird/proxy/internal/middleware/bodytap"
+	mwbuiltin "github.com/netbirdio/netbird/proxy/internal/middleware/builtin"
+	_ "github.com/netbirdio/netbird/proxy/internal/middleware/builtin/cost_meter"
+	_ "github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_guardrail"
+	_ "github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_identity_inject"
+	_ "github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_limit_check"
+	_ "github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_limit_record"
+	_ "github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_request_parser"
+	_ "github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_response_parser"
+	_ "github.com/netbirdio/netbird/proxy/internal/middleware/builtin/llm_router"
+	"github.com/netbirdio/netbird/proxy/internal/proxy"
+	nbproxytypes "github.com/netbirdio/netbird/proxy/internal/types"
+	"github.com/netbirdio/netbird/shared/management/proto"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// TestReverseProxy_VertexGuardrail_ModelAllowlist drives Anthropic-on-Vertex
+// requests (model in the URL path, not the body — the shape Claude Code sends)
+// through the full synthesized middleware chain against an in-process
+// management stack and a fake Vertex upstream. It reproduces the customer
+// report end to end: with a guardrail allowlisting only Sonnet, an Opus request
+// must be denied by the guardrail (model_blocked) before it reaches the
+// upstream, and the allowed Sonnet request must reach it.
+//
+// Two provider shapes are covered because they exercise different code:
+//
+//   - "catch_all": provider registered with no models (the historical Vertex
+//     default). The router waves every model through, so the guardrail is the
+//     only thing that can block Opus — the direct regression guard for "the
+//     disallowed model is never blocked".
+//   - "versioned_models": provider registered with the raw "@version" model ids
+//     the Google console documents. The router must normalize the registered id
+//     to route the version-stripped request model (otherwise even Sonnet denies
+//     as model_not_routable), and the guardrail must still block Opus.
+//
+// The upstream-hit counter proves whether a denied request actually
+// short-circuited or leaked through to Google.
+func TestReverseProxy_VertexGuardrail_ModelAllowlist(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("sqlite store not supported on Windows")
+	}
+
+	cases := []struct {
+		name           string
+		providerModels []agentNetworkTypes.ProviderModel
+	}{
+		{name: "catch_all", providerModels: nil},
+		{name: "versioned_models", providerModels: []agentNetworkTypes.ProviderModel{
+			{ID: "claude-sonnet-4-5@20250929"},
+			{ID: "claude-opus-4-6@20250514"},
+		}},
+	}
+	for _, tc := range cases {
+		tc := tc
+		t.Run(tc.name, func(t *testing.T) {
+			runVertexGuardrailCase(t, tc.providerModels)
+		})
+	}
+}
+
+func runVertexGuardrailCase(t *testing.T, providerModels []agentNetworkTypes.ProviderModel) {
+	t.Helper()
+
+	const (
+		testAccountID = "acct-vertex-guard-1"
+		testAdminUser = "user-admin-1"
+		adminGroupID  = "grp-admins"
+		providerID    = "prov-vertex-test"
+		guardrailID   = "ainguard-sonnet-only"
+		cluster       = "test.proxy.local"
+		subdomain     = "vertexguard"
+	)
+	testLogger := log.New()
+	testLogger.SetLevel(log.PanicLevel)
+
+	ctx := context.Background()
+
+	// Fake Vertex upstream: any request that reaches it is a guardrail miss for
+	// the disallowed model.
+	var upstreamHits atomic.Int64
+	upstreamBody := []byte(`{"id":"msg_x","type":"message","role":"assistant","model":"claude","content":[{"type":"text","text":"pong"}],"usage":{"input_tokens":5,"output_tokens":2}}`)
+	upstream := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		upstreamHits.Add(1)
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write(upstreamBody)
+	}))
+	t.Cleanup(upstream.Close)
+
+	// In-process management gRPC (bufconn) over a real sqlite store + manager.
+	st, cleanup, err := store.NewTestStoreFromSQL(ctx, "", t.TempDir())
+	require.NoError(t, err, "real sqlite test store must come up")
+	t.Cleanup(cleanup)
+
+	anMgr := agentnetwork.NewManager(st, nil, nil, nil)
+	server := &mgmtgrpc.ProxyServiceServer{}
+	server.SetAgentNetworkLimitsService(anMgr)
+
+	lis := bufconn.Listen(1024 * 1024)
+	srv := grpc.NewServer()
+	proto.RegisterProxyServiceServer(srv, server)
+	go func() { _ = srv.Serve(lis) }()
+	t.Cleanup(srv.Stop)
+
+	conn, err := grpc.NewClient("passthrough:///bufnet",
+		grpc.WithContextDialer(func(_ context.Context, _ string) (net.Conn, error) { return lis.Dial() }),
+		grpc.WithTransportCredentials(insecure.NewCredentials()),
+	)
+	require.NoError(t, err)
+	t.Cleanup(func() { _ = conn.Close() })
+	mgmtClient := proto.NewProxyServiceClient(conn)
+
+	require.NoError(t, st.SaveAgentNetworkSettings(ctx, &agentNetworkTypes.Settings{
+		AccountID: testAccountID,
+		Cluster:   cluster,
+		Subdomain: subdomain,
+	}))
+	require.NoError(t, st.SaveAgentNetworkProvider(ctx, &agentNetworkTypes.Provider{
+		ID:          providerID,
+		AccountID:   testAccountID,
+		ProviderID:  "vertex_ai_api",
+		Name:        "vertex-guard-test",
+		UpstreamURL: upstream.URL,
+		// A static bearer (not "keyfile::…") so the router injects a static auth
+		// header instead of minting a GCP OAuth token at request time — the mint
+		// needs network egress and would deny before the guardrail runs, masking
+		// the guardrail decision this test exists to observe.
+		APIKey:            "static-vertex-token",
+		Enabled:           true,
+		Models:            providerModels,
+		SessionPrivateKey: "priv",
+		SessionPublicKey:  "pub",
+	}))
+	// Guardrail allowlisting ONLY Sonnet.
+	require.NoError(t, st.SaveAgentNetworkGuardrail(ctx, &agentNetworkTypes.Guardrail{
+		ID:        guardrailID,
+		AccountID: testAccountID,
+		Name:      "sonnet-only",
+		Checks: agentNetworkTypes.GuardrailChecks{
+			ModelAllowlist: agentNetworkTypes.GuardrailModelAllowlist{
+				Enabled: true,
+				Models:  []string{"claude-sonnet-4-5"},
+			},
+		},
+	}))
+	require.NoError(t, st.SaveAgentNetworkPolicy(ctx, &agentNetworkTypes.Policy{
+		ID:                     "ainpol-vertex-guard",
+		AccountID:              testAccountID,
+		Name:                   "admins-vertex",
+		Enabled:                true,
+		SourceGroups:           []string{adminGroupID},
+		DestinationProviderIDs: []string{providerID},
+		GuardrailIDs:           []string{guardrailID},
+	}))
+
+	services, err := agentnetwork.SynthesizeServices(ctx, st, testAccountID)
+	require.NoError(t, err)
+	require.Len(t, services, 1, "exactly one synth service expected")
+	synthSvc := services[0]
+	require.NotEmpty(t, synthSvc.Targets, "synth target must exist")
+
+	mwbuiltin.Configure(ctx, t.TempDir(), nil, testLogger, mgmtClient)
+	registry := mwbuiltin.DefaultRegistry()
+	mwMetrics, err := middleware.NewMetrics(nil)
+	require.NoError(t, err)
+	mwMgr := middleware.NewManager(0, mwMetrics, testLogger)
+	mwMgr.SetResolver(middleware.NewResolver(registry))
+
+	specs := make([]middleware.Spec, 0, len(synthSvc.Targets[0].Options.Middlewares))
+	for _, mw := range synthSvc.Targets[0].Options.Middlewares {
+		var slot middleware.Slot
+		switch mw.Slot {
+		case rpservice.MiddlewareSlotOnRequest:
+			slot = middleware.SlotOnRequest
+		case rpservice.MiddlewareSlotOnResponse:
+			slot = middleware.SlotOnResponse
+		case rpservice.MiddlewareSlotTerminal:
+			slot = middleware.SlotTerminal
+		default:
+			t.Fatalf("unknown middleware slot %q on %s", mw.Slot, mw.ID)
+		}
+		specs = append(specs, middleware.Spec{
+			ID:        mw.ID,
+			Slot:      slot,
+			Enabled:   mw.Enabled,
+			FailMode:  middleware.FailOpen,
+			Timeout:   middleware.DefaultTimeout,
+			RawConfig: append([]byte(nil), mw.ConfigJSON...),
+			CanMutate: mw.CanMutate,
+		})
+	}
+
+	serviceIDStr := synthSvc.ID
+	require.NoError(t, mwMgr.Rebuild(serviceIDStr, []middleware.PathTargetBinding{{
+		ServiceID: serviceIDStr,
+		PathID:    "/",
+		Specs:     specs,
+	}}))
+
+	upstreamURL, err := url.Parse(upstream.URL)
+	require.NoError(t, err)
+
+	rp := proxy.NewReverseProxy(http.DefaultTransport, "auto", nil, testLogger, proxy.WithMiddlewareManager(mwMgr))
+	rp.AddMapping(proxy.Mapping{
+		ID:        nbproxytypes.ServiceID(serviceIDStr),
+		AccountID: nbproxytypes.AccountID(testAccountID),
+		Host:      synthSvc.Domain,
+		Paths: map[string]*proxy.PathTarget{
+			"/": {
+				URL:            upstreamURL,
+				DirectUpstream: true,
+				AgentNetwork:   true,
+				Middlewares:    specs,
+				CaptureConfig: &bodytap.Config{
+					MaxRequestBytes:  1 << 20,
+					MaxResponseBytes: 1 << 20,
+					ContentTypes:     []string{"application/json", "text/event-stream"},
+				},
+			},
+		},
+	})
+
+	// vertexBody carries no "model" field — the model lives in the URL path,
+	// exactly as an Anthropic SDK client (Claude Code) posts to Vertex.
+	const vertexBody = `{"anthropic_version":"vertex-2023-10-16","max_tokens":64,"messages":[{"role":"user","content":"Reply with exactly: pong"}]}`
+
+	send := func(t *testing.T, model string) (int, int64, string) {
+		t.Helper()
+		before := upstreamHits.Load()
+		path := "/v1/projects/corp-gcp-it-all-claude/locations/global/publishers/anthropic/models/" + model + ":rawPredict"
+		req := httptest.NewRequest("POST", "https://"+synthSvc.Domain+path, strings.NewReader(vertexBody))
+		req.Host = synthSvc.Domain
+		req.Header.Set("Content-Type", "application/json")
+
+		cd := proxy.NewCapturedData("req-" + model)
+		cd.SetServiceID(nbproxytypes.ServiceID(serviceIDStr))
+		cd.SetAccountID(nbproxytypes.AccountID(testAccountID))
+		cd.SetUserID(testAdminUser)
+		cd.SetUserGroups([]string{adminGroupID})
+		cd.SetAuthMethod("tunnel_peer")
+		req = req.WithContext(proxy.WithCapturedData(req.Context(), cd))
+
+		w := httptest.NewRecorder()
+		rp.ServeHTTP(w, req)
+		return w.Code, upstreamHits.Load() - before, w.Body.String()
+	}
+
+	// The disallowed model (Opus) MUST be denied by the guardrail (model_blocked)
+	// before the upstream — this is the customer-reported bug (Opus let through
+	// under a Sonnet-only allowlist). The deny code distinguishes a guardrail
+	// block from a router model_not_routable, which would deny for the wrong
+	// reason.
+	t.Run("opus_denied_by_guardrail", func(t *testing.T) {
+		code, hits, body := send(t, "claude-opus-4-6")
+		assert.Equal(t, http.StatusForbidden, code, "Opus must be denied under a Sonnet-only allowlist; body=%s", body)
+		assert.Contains(t, body, "llm_policy.model_blocked", "denial must come from the guardrail allowlist, not routing; body=%s", body)
+		assert.Equal(t, int64(0), hits, "a denied request must never reach the Vertex upstream")
+	})
+
+	// The allowed model (Sonnet) passes the guardrail and reaches the upstream.
+	t.Run("sonnet_allowed", func(t *testing.T) {
+		code, hits, body := send(t, "claude-sonnet-4-5")
+		assert.Equal(t, http.StatusOK, code, "Sonnet is allowlisted and must be served; body=%s", body)
+		assert.Equal(t, int64(1), hits, "the allowed request must reach the Vertex upstream exactly once")
+	})
+}

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
-}

proxy/trustedproxy_test.go

@@ -1,90 +0,0 @@
-package proxy
-
-import (
-	"net/netip"
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-func TestParseTrustedProxies(t *testing.T) {
-	tests := []struct {
-		name    string
-		raw     string
-		want    []netip.Prefix
-		wantErr bool
-	}{
-		{
-			name: "empty string returns nil",
-			raw:  "",
-			want: nil,
-		},
-		{
-			name: "single CIDR",
-			raw:  "10.0.0.0/8",
-			want: []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")},
-		},
-		{
-			name: "single bare IPv4",
-			raw:  "1.2.3.4",
-			want: []netip.Prefix{netip.MustParsePrefix("1.2.3.4/32")},
-		},
-		{
-			name: "single bare IPv6",
-			raw:  "::1",
-			want: []netip.Prefix{netip.MustParsePrefix("::1/128")},
-		},
-		{
-			name: "comma-separated CIDRs",
-			raw:  "10.0.0.0/8, 192.168.1.0/24",
-			want: []netip.Prefix{
-				netip.MustParsePrefix("10.0.0.0/8"),
-				netip.MustParsePrefix("192.168.1.0/24"),
-			},
-		},
-		{
-			name: "mixed CIDRs and bare IPs",
-			raw:  "10.0.0.0/8, 1.2.3.4, fd00::/8",
-			want: []netip.Prefix{
-				netip.MustParsePrefix("10.0.0.0/8"),
-				netip.MustParsePrefix("1.2.3.4/32"),
-				netip.MustParsePrefix("fd00::/8"),
-			},
-		},
-		{
-			name: "whitespace around entries",
-			raw:  "  10.0.0.0/8 , 192.168.0.0/16  ",
-			want: []netip.Prefix{
-				netip.MustParsePrefix("10.0.0.0/8"),
-				netip.MustParsePrefix("192.168.0.0/16"),
-			},
-		},
-		{
-			name: "trailing comma produces no extra entry",
-			raw:  "10.0.0.0/8,",
-			want: []netip.Prefix{netip.MustParsePrefix("10.0.0.0/8")},
-		},
-		{
-			name:    "invalid entry",
-			raw:     "not-an-ip",
-			wantErr: true,
-		},
-		{
-			name:    "partially invalid",
-			raw:     "10.0.0.0/8, garbage",
-			wantErr: true,
-		},
-	}
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			got, err := ParseTrustedProxies(tt.raw)
-			if tt.wantErr {
-				require.Error(t, err)
-				return
-			}
-			require.NoError(t, err)
-			assert.Equal(t, tt.want, got)
-		})
-	}
-}

relay/cmd/root.go

@@ -24,6 +24,7 @@ import (
 	"github.com/netbirdio/netbird/shared/metrics"
 	"github.com/netbirdio/netbird/shared/relay/auth"
 	"github.com/netbirdio/netbird/stun"
+	"github.com/netbirdio/netbird/trustedproxy"
 	"github.com/netbirdio/netbird/util"
 )
 
@@ -45,6 +46,9 @@ type Config struct {
 	LogLevel                 string
 	LogFile                  string
 	HealthcheckListenAddress string
+	// TrustedProxies is a comma-separated list of upstream proxy CIDRs/IPs whose
+	// X-Real-Ip/X-Real-Port headers are trusted. Empty means never trust these headers.
+	TrustedProxies string
 	// STUN server configuration
 	EnableSTUN   bool
 	STUNPorts    []int
@@ -116,6 +120,7 @@ func init() {
 	rootCmd.PersistentFlags().StringVar(&cobraConfig.LogLevel, "log-level", "info", "log level")
 	rootCmd.PersistentFlags().StringVar(&cobraConfig.LogFile, "log-file", "console", "log file")
 	rootCmd.PersistentFlags().StringVarP(&cobraConfig.HealthcheckListenAddress, "health-listen-address", "H", ":9000", "listen address of healthcheck server")
+	rootCmd.PersistentFlags().StringVar(&cobraConfig.TrustedProxies, "trusted-proxies", "", "comma-separated list of upstream proxy CIDRs or IPs whose X-Real-Ip/X-Real-Port headers are trusted; leave empty to always use the direct connection address")
 	rootCmd.PersistentFlags().BoolVar(&cobraConfig.EnableSTUN, "enable-stun", false, "enable embedded STUN server")
 	rootCmd.PersistentFlags().IntSliceVar(&cobraConfig.STUNPorts, "stun-ports", []int{3478}, "ports for the embedded STUN server (can be specified multiple times or comma-separated)")
 	rootCmd.PersistentFlags().StringVar(&cobraConfig.STUNLogLevel, "stun-log-level", "info", "log level for STUN server (panic, fatal, error, warn, info, debug, trace)")
@@ -155,8 +160,15 @@ func execute(cmd *cobra.Command, args []string) error {
 		return fmt.Errorf("setup metrics: %v", err)
 	}
 
+	trustedProxies, err := trustedproxy.Parse(cobraConfig.TrustedProxies)
+	if err != nil {
+		log.Debugf("failed to parse trusted proxies: %s", err)
+		return fmt.Errorf("failed to parse trusted proxies: %s", err)
+	}
+
 	srvListenerCfg := server.ListenerConfig{
-		Address: cobraConfig.ListenAddress,
+		Address:        cobraConfig.ListenAddress,
+		TrustedProxies: trustedProxies,
 	}
 
 	tlsConfig, tlsSupport, err := handleTLSConfig(cobraConfig)
@@ -173,7 +185,7 @@ func execute(cmd *cobra.Command, args []string) error {
 	}
 
 	hashedSecret := sha256.Sum256([]byte(cobraConfig.AuthSecret))
-	authenticator := auth.NewTimedHMACValidator(hashedSecret[:], 24*time.Hour)
+	authenticator := auth.NewTimedHMACValidator(hashedSecret[:])
 
 	cfg := server.Config{
 		Meter:          metricsServer.Meter,

relay/server/handshake.go

@@ -5,14 +5,8 @@ import (
 	"fmt"
 	"time"
 
-	log "github.com/sirupsen/logrus"
-
 	"github.com/netbirdio/netbird/relay/server/listener"
 	"github.com/netbirdio/netbird/shared/relay/messages"
-	//nolint:staticcheck
-	"github.com/netbirdio/netbird/shared/relay/messages/address"
-	//nolint:staticcheck
-	authmsg "github.com/netbirdio/netbird/shared/relay/messages/auth"
 )
 
 const (
@@ -23,55 +17,30 @@ const (
 
 type Validator interface {
 	Validate(any) error
-	// Deprecated: Use Validate instead.
-	ValidateHelloMsgType(any) error
 }
 
-// preparedMsg contains the marshalled success response messages
+// preparedMsg contains the marshalled success response message
 type preparedMsg struct {
-	responseHelloMsg []byte
-	responseAuthMsg  []byte
+	responseAuthMsg []byte
 }
 
 func newPreparedMsg(instanceURL string) (*preparedMsg, error) {
-	rhm, err := marshalResponseHelloMsg(instanceURL)
-	if err != nil {
-		return nil, err
-	}
-
 	ram, err := messages.MarshalAuthResponse(instanceURL)
 	if err != nil {
 		return nil, fmt.Errorf("failed to marshal auth response msg: %w", err)
 	}
 
 	return &preparedMsg{
-		responseHelloMsg: rhm,
-		responseAuthMsg:  ram,
+		responseAuthMsg: ram,
 	}, nil
 }
 
-func marshalResponseHelloMsg(instanceURL string) ([]byte, error) {
-	addr := &address.Address{URL: instanceURL}
-	addrData, err := addr.Marshal()
-	if err != nil {
-		return nil, fmt.Errorf("failed to marshal response address: %w", err)
-	}
-
-	//nolint:staticcheck
-	responseMsg, err := messages.MarshalHelloResponse(addrData)
-	if err != nil {
-		return nil, fmt.Errorf("failed to marshal hello response: %w", err)
-	}
-	return responseMsg, nil
-}
-
 type handshake struct {
 	conn        listener.Conn
 	validator   Validator
 	preparedMsg *preparedMsg
 
-	handshakeMethodAuth bool
-	peerID              *messages.PeerID
+	peerID *messages.PeerID
 }
 
 func (h *handshake) handshakeReceive(ctx context.Context) (*messages.PeerID, error) {
@@ -93,17 +62,11 @@ func (h *handshake) handshakeReceive(ctx context.Context) (*messages.PeerID, err
 		return nil, fmt.Errorf("determine message type from %s: %w", h.conn.RemoteAddr(), err)
 	}
 
-	var peerID *messages.PeerID
-	switch msgType {
-	//nolint:staticcheck
-	case messages.MsgTypeHello:
-		peerID, err = h.handleHelloMsg(buf)
-	case messages.MsgTypeAuth:
-		h.handshakeMethodAuth = true
-		peerID, err = h.handleAuthMsg(buf)
-	default:
+	if msgType != messages.MsgTypeAuth {
 		return nil, fmt.Errorf("invalid message type %d from %s", msgType, h.conn.RemoteAddr())
 	}
+
+	peerID, err := h.handleAuthMsg(buf)
 	if err != nil {
 		return peerID, err
 	}
@@ -112,46 +75,17 @@ func (h *handshake) handshakeReceive(ctx context.Context) (*messages.PeerID, err
 }
 
 func (h *handshake) handshakeResponse(ctx context.Context) error {
-	var responseMsg []byte
-	if h.handshakeMethodAuth {
-		responseMsg = h.preparedMsg.responseAuthMsg
-	} else {
-		responseMsg = h.preparedMsg.responseHelloMsg
-	}
-
-	if _, err := h.conn.Write(ctx, responseMsg); err != nil {
+	if _, err := h.conn.Write(ctx, h.preparedMsg.responseAuthMsg); err != nil {
 		return fmt.Errorf("handshake response write to %s (%s): %w", h.peerID, h.conn.RemoteAddr(), err)
 	}
 
 	return nil
 }
 
-func (h *handshake) handleHelloMsg(buf []byte) (*messages.PeerID, error) {
-	//nolint:staticcheck
-	peerID, authData, err := messages.UnmarshalHelloMsg(buf)
-	if err != nil {
-		return nil, fmt.Errorf("unmarshal hello message: %w", err)
-	}
-
-	log.Warnf("peer %s (%s) is using deprecated initial message type", peerID, h.conn.RemoteAddr())
-
-	authMsg, err := authmsg.UnmarshalMsg(authData)
-	if err != nil {
-		return nil, fmt.Errorf("unmarshal auth message: %w", err)
-	}
-
-	//nolint:staticcheck
-	if err := h.validator.ValidateHelloMsgType(authMsg.AdditionalData); err != nil {
-		return nil, fmt.Errorf("validate %s (%s): %w", peerID, h.conn.RemoteAddr(), err)
-	}
-
-	return peerID, nil
-}
-
 func (h *handshake) handleAuthMsg(buf []byte) (*messages.PeerID, error) {
 	rawPeerID, authPayload, err := messages.UnmarshalAuthMsg(buf)
 	if err != nil {
-		return nil, fmt.Errorf("unmarshal hello message: %w", err)
+		return nil, fmt.Errorf("unmarshal auth message: %w", err)
 	}
 
 	if err := h.validator.Validate(authPayload); err != nil {

relay/server/listener/quic/listener.go

@@ -51,7 +51,10 @@ func (l *Listener) Listen(acceptFn func(conn relaylistener.Conn)) error {
 
 		log.Infof("QUIC client connected from: %s", session.RemoteAddr())
 		conn := NewConn(session)
-		acceptFn(conn)
+		// Run the accept handler (which performs the pre-auth handshake) in its
+		// own goroutine so a slow or stalled handshake cannot block accepting
+		// further connections.
+		go acceptFn(conn)
 	}
 }
 

relay/server/listener/ws/listener.go

@@ -15,6 +15,7 @@ import (
 	"github.com/netbirdio/netbird/relay/protocol"
 	relaylistener "github.com/netbirdio/netbird/relay/server/listener"
 	"github.com/netbirdio/netbird/shared/relay"
+	"github.com/netbirdio/netbird/trustedproxy"
 )
 
 const (
@@ -27,6 +28,9 @@ type Listener struct {
 	Address string
 	// TLSConfig is the TLS configuration for the server.
 	TLSConfig *tls.Config
+	// TrustedProxies is the set of upstream proxies whose X-Real-Ip/X-Real-Port
+	// headers are trusted. Headers from any other immediate peer are ignored.
+	TrustedProxies *trustedproxy.List
 
 	server   *http.Server
 	acceptFn func(conn relaylistener.Conn)
@@ -75,7 +79,7 @@ func (l *Listener) Shutdown(ctx context.Context) error {
 }
 
 func (l *Listener) onAccept(w http.ResponseWriter, r *http.Request) {
-	connRemoteAddr := remoteAddr(r)
+	connRemoteAddr := remoteAddr(r, l.TrustedProxies)
 
 	acceptOptions := &websocket.AcceptOptions{
 		OriginPatterns: []string{"*"},
@@ -102,9 +106,17 @@ func (l *Listener) onAccept(w http.ResponseWriter, r *http.Request) {
 	l.acceptFn(conn)
 }
 
-func remoteAddr(r *http.Request) string {
-	if r.Header.Get("X-Real-Ip") == "" || r.Header.Get("X-Real-Port") == "" {
+func remoteAddr(r *http.Request, trustedProxies *trustedproxy.List) string {
+	realIP := r.Header.Get("X-Real-Ip")
+	realPort := r.Header.Get("X-Real-Port")
+	if realIP == "" || realPort == "" {
 		return r.RemoteAddr
 	}
-	return net.JoinHostPort(r.Header.Get("X-Real-Ip"), r.Header.Get("X-Real-Port"))
+
+	if !trustedProxies.IsTrusted(r.RemoteAddr) {
+		log.Debugf("ignoring X-Real-Ip header from untrusted peer %s", r.RemoteAddr)
+		return r.RemoteAddr
+	}
+
+	return net.JoinHostPort(realIP, realPort)
 }