Assert the full flow order in both directions

Prepares the repository for the release-branch process agreed internally:
one long-lived release-0.N branch per minor, with fixes backported by
cherry-pick and patch releases tagged from the branch.

Pushes to release-* branches now run the Release workflow and publish
immutable sha-* container images, the way pushes to main already do, so
a release branch can be tested before it is tagged. Release branches
never publish the floating "main" image tag. The push-to-main CI
workflows (Go tests on all platforms, frontend UI, install script,
mobile/wasm validation, infrastructure files, license check) also run
on release-* pushes; pull request triggers were already unfiltered, so
backport PRs were covered — this closes the post-merge gap.

Releases are no longer marked latest before signing: make_latest is
now false in all four goreleaser configs, so a release stays published
but not latest until the signing pipeline uploads the signed Windows
and macOS artifacts and marks it latest itself. Previously the release
became GitHub's "Latest release" at publish time, and the download
endpoints that resolve through the latest-release API could serve a
release whose signed installers did not exist yet. prerelease: auto
additionally labels rc tags as prereleases, so a release candidate can
never take the latest slot.

The trigger_sync_tag job is removed: it dispatched a downstream
image build on every v* tag (release candidates included), which would
race the deliberate release-branch build on every release. The android
and ios submodule bumps are unchanged.

Also sets perennial-regex = "^release-" so git-town never syncs or
ships a release branch into main.

.git-branches.toml

@@ -3,7 +3,7 @@
 [branches]
 main = "main"
 perennials = []
-perennial-regex = ""
+perennial-regex = "^release-"
 
 [create]
 new-branch-type = "feature"

.github/workflows/check-license-dependencies.yml

@@ -2,7 +2,7 @@ name: Check License Dependencies
 
 on:
   push:
-    branches: [main]
+    branches: [main, "release-*"]
     paths:
       - "go.mod"
       - "go.sum"

.github/workflows/frontend-ui.yml

@@ -10,6 +10,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
     paths:
       - "client/ui/frontend/**"
       - "client/ui/i18n/**"

.github/workflows/golang-test-darwin.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
   pull_request:
 
 concurrency:

.github/workflows/golang-test-freebsd.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
   pull_request:
 
 concurrency:

.github/workflows/golang-test-linux.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
   pull_request:
 
 concurrency:

.github/workflows/golang-test-windows.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
   pull_request:
 
 env:

.github/workflows/install-script-test.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
   pull_request:
     paths:
       - "release_files/install.sh"

.github/workflows/mobile-build-validation.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
   pull_request:
 
 concurrency:

.github/workflows/release.yml

@@ -6,6 +6,7 @@ on:
       - "v*"
     branches:
       - main
+      - "release-*"
   pull_request:
 
 env:
@@ -254,15 +255,23 @@ jobs:
         id: tag_and_push_images
         if: |
           (github.event_name == 'pull_request' && github.event.pull_request.head.repo.full_name == github.repository) ||
-          (github.event_name == 'push' && github.ref == 'refs/heads/main')
+          (github.event_name == 'push' && (github.ref == 'refs/heads/main' || startsWith(github.ref, 'refs/heads/release-')))
         run: |
           set -euo pipefail
 
+          # $GITHUB_REF / $GITHUB_EVENT_NAME are read from the runner
+          # environment rather than substituted into this script with the
+          # workflow expression syntax: branch names may legally contain
+          # $(…), and interpolating github.ref would execute it.
           resolve_tags() {
-            if [[ "${{ github.event_name }}" == "pull_request" ]]; then
+            if [[ "$GITHUB_EVENT_NAME" == "pull_request" ]]; then
               echo "pr-${{ github.event.pull_request.number }}"
-            else
+            elif [[ "$GITHUB_REF" == "refs/heads/main" ]]; then
               echo "main sha-$(git rev-parse --short HEAD)"
+            else
+              # Release branches get an immutable sha-* tag only — the floating
+              # "main" tag must never move from a release branch.
+              echo "sha-$(git rev-parse --short HEAD)"
             fi
           }
 

.github/workflows/sync-tag.yml

@@ -9,21 +9,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.actor_id }}
   cancel-in-progress: true
 
-# Receiving workflows (cloud sync-tag, mobile bump-netbird) expect the short
-# tag form (e.g. v0.30.0), not refs/tags/v0.30.0 — github.ref_name, not github.ref.
+# The receiving bump-netbird workflows expect the short tag form
+# (e.g. v0.30.0), not refs/tags/v0.30.0 — github.ref_name, not github.ref.
 jobs:
-  trigger_sync_tag:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Trigger release tag sync
-        uses: benc-uk/workflow-dispatch@31e2b3319479a63f0ab15bf800eff9e913504e26 # v1.3.2
-        with:
-          workflow: sync-tag.yml
-          ref: main
-          repo: ${{ secrets.UPSTREAM_REPO }}
-          token: ${{ secrets.NC_GITHUB_TOKEN }}
-          inputs: '{ "tag": "${{ github.ref_name }}" }'
-
   trigger_android_bump:
     runs-on: ubuntu-latest
     if: github.event.created && !github.event.deleted && startsWith(github.ref, 'refs/tags/v') && !contains(github.ref_name, '-')

.github/workflows/test-infrastructure-files.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
   pull_request:
     paths:
       - "infrastructure_files/**"

.github/workflows/wasm-build-validation.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - main
+      - "release-*"
   pull_request:
 
 concurrency:

.goreleaser.yaml

@@ -468,6 +468,13 @@ checksum:
     - glob: ./infrastructure_files/migrate-to-enterprise.sh
 
 release:
+  # The signing pipeline (netbirdio/sign-pipelines, dispatched by
+  # trigger_signer) marks the release latest once the Windows and macOS
+  # artifacts are signed. Without this override goreleaser marks it latest
+  # at publish time, while those artifacts are still unsigned.
+  make_latest: false
+  # Mark x.y.z-rc.* and other prerelease tags as prereleases on GitHub.
+  prerelease: auto
   extra_files:
     - glob: ./infrastructure_files/getting-started-with-zitadel.sh
     - glob: ./release_files/install.sh

.goreleaser_ui.yaml

@@ -96,6 +96,7 @@ nfpms:
       - netbird (>= 0.75.0)
       - libgtk-4-1 (>= 4.14)
       - libwebkitgtk-6.0-4
+      - xdg-utils
 
   - maintainer: Netbird <dev@netbird.io>
     description: Netbird client UI.
@@ -119,6 +120,7 @@ nfpms:
       - netbird >= 0.75.0
       - (gtk4 >= 4.14 or libgtk-4-1 >= 4.14)
       - (webkitgtk6.0 or libwebkitgtk-6_0-4)
+      - xdg-utils
 
     rpm:
       signature:
@@ -142,3 +144,11 @@ uploads:
     target: https://pkgs.wiretrustee.com/yum/{{ .Arch }}{{ if .Arm }}{{ .Arm }}{{ end }}
     username: dev@wiretrustee.com
     method: PUT
+
+release:
+  # Uploads into the release created by the main .goreleaser.yaml run.
+  # make_latest stays false everywhere: the signing pipeline
+  # (netbirdio/sign-pipelines) marks the release latest after the Windows
+  # and macOS artifacts are signed.
+  make_latest: false
+  prerelease: auto

.goreleaser_ui_darwin.yaml

@@ -43,3 +43,11 @@ checksum:
   name_template: "{{ .ProjectName }}_darwin_checksums.txt"
 changelog:
   disable: true
+
+release:
+  # Uploads into the release created by the main .goreleaser.yaml run.
+  # make_latest stays false everywhere: the signing pipeline
+  # (netbirdio/sign-pipelines) marks the release latest after the Windows
+  # and macOS artifacts are signed.
+  make_latest: false
+  prerelease: auto

.goreleaser_ui_gtk3.yaml

@@ -43,19 +43,17 @@ archives:
       - netbird-ui-gtk3
 
 nfpms:
-  # Same package_name as the GTK4 packages -- the two are mutually-exclusive
-  # alternatives served from separate repo paths (see uploads below); a given
-  # distro points at exactly one of them. The file names must still differ:
-  # the Debian pool is shared storage keyed by file name, so a default-named
-  # gtk3 .deb would overwrite the stable one.
+  # Mutually-exclusive alternative to the GTK4 netbird-ui package -- both
+  # ship the same /usr/bin/netbird-ui from the shared stable/yum repos, so
+  # this one carries its own name and conflicts with the GTK4 package.
   - maintainer: Netbird <dev@netbird.io>
     description: Netbird client UI.
     homepage: https://netbird.io/
     license: BSD-3-Clause
     vendor: NetBird
     id: netbird_ui_deb_gtk3
-    package_name: netbird-ui
-    file_name_template: "{{ .PackageName }}-gtk3_{{ .Version }}_{{ .Os }}_{{ .Arch }}"
+    package_name: netbird-ui-gtk3
+    file_name_template: "{{ .PackageName }}_{{ .Version }}_{{ .Os }}_{{ .Arch }}"
     builds:
       - netbird-ui-gtk3
     formats:
@@ -67,10 +65,15 @@ nfpms:
         dst: /usr/share/applications/org.wails.netbird.desktop
       - src: client/ui/build/appicon.png
         dst: /usr/share/pixmaps/netbird.png
+    conflicts:
+      - netbird-ui
+    replaces:
+      - netbird-ui
     dependencies:
       - netbird (>= 0.75.0)
       - libgtk-3-0
       - libwebkit2gtk-4.1-0
+      - xdg-utils
 
   - maintainer: Netbird <dev@netbird.io>
     description: Netbird client UI.
@@ -78,8 +81,8 @@ nfpms:
     license: BSD-3-Clause
     vendor: NetBird
     id: netbird_ui_rpm_gtk3
-    package_name: netbird-ui
-    file_name_template: "{{ .PackageName }}-gtk3_{{ .Version }}_{{ .Os }}_{{ .Arch }}"
+    package_name: netbird-ui-gtk3
+    file_name_template: "{{ .PackageName }}_{{ .Version }}_{{ .Os }}_{{ .Arch }}"
     builds:
       - netbird-ui-gtk3
     formats:
@@ -91,10 +94,15 @@ nfpms:
         dst: /usr/share/applications/org.wails.netbird.desktop
       - src: client/ui/build/appicon.png
         dst: /usr/share/pixmaps/netbird.png
+    # No `replaces` here: nfpm maps it to rpm Obsoletes, which would make
+    # dnf swap installed GTK4 netbird-ui packages for this one on upgrade.
+    conflicts:
+      - netbird-ui
     dependencies:
       - netbird >= 0.75.0
       - (gtk3 or libgtk-3-0)
       - (webkit2gtk4.1 or libwebkit2gtk-4_1-0)
+      - xdg-utils
 
     rpm:
       signature:
@@ -109,23 +117,28 @@ changelog:
   disable: true
 
 uploads:
-  # The gtk3 packages reuse the netbird-ui package name, so they live in
-  # dedicated repo paths (deb distribution `gtk3`, yum path `yum-gtk3`) that
-  # legacy distros point their repo config at.
-  - name: debian-gtk3
+  - name: debian
     skip: "{{ .Env.SKIP_PUBLISH }}"
     ids:
       - netbird_ui_deb_gtk3
     mode: archive
-    target: https://pkgs.wiretrustee.com/debian/pool/{{ .ArtifactName }};deb.distribution=gtk3;deb.component=main;deb.architecture={{ if .Arm }}armhf{{ else }}{{ .Arch }}{{ end }};deb.package=
+    target: https://pkgs.wiretrustee.com/debian/pool/{{ .ArtifactName }};deb.distribution=stable;deb.component=main;deb.architecture={{ if .Arm }}armhf{{ else }}{{ .Arch }}{{ end }};deb.package=
     username: dev@wiretrustee.com
     method: PUT
 
-  - name: yum-gtk3
+  - name: yum
     skip: "{{ .Env.SKIP_PUBLISH }}"
     ids:
       - netbird_ui_rpm_gtk3
     mode: archive
-    target: https://pkgs.wiretrustee.com/yum-gtk3/{{ .Arch }}{{ if .Arm }}{{ .Arm }}{{ end }}
+    target: https://pkgs.wiretrustee.com/yum/{{ .Arch }}{{ if .Arm }}{{ .Arm }}{{ end }}
     username: dev@wiretrustee.com
     method: PUT
+
+release:
+  # Uploads into the release created by the main .goreleaser.yaml run.
+  # make_latest stays false everywhere: the signing pipeline
+  # (netbirdio/sign-pipelines) marks the release latest after the Windows
+  # and macOS artifacts are signed.
+  make_latest: false
+  prerelease: auto

CONTRIBUTING.md

@@ -112,6 +112,7 @@ aligns with our security standards and design expectations.
         - [Test suite](#test-suite)
     - [Checklist before submitting a PR](#checklist-before-submitting-a-pr)
     - [When we close a PR](#when-we-close-a-pr)
+    - [Translations](#translations)
     - [Other project repositories](#other-project-repositories)
     - [Contributor License Agreement](#contributor-license-agreement)
 
@@ -612,6 +613,17 @@ A closed PR is not a rejected idea. Take it back to the
 [discussion](https://github.com/netbirdio/netbird/discussions), settle the
 approach, and reopen the work from there.
 
+## Translations
+
+Desktop UI translations are not contributed through pull requests. Translate on
+[Crowdin](https://crowdin.com/project/netbird) instead: no ticket needed, just
+join the project and pick your language. Crowdin syncs with this repository and
+opens the service PRs itself, so hand-edited locale files would conflict with
+the next sync. Style, terminology, and review guidance live in
+[client/ui/i18n/TRANSLATING.md](client/ui/i18n/TRANSLATING.md). To request a
+language the project does not offer yet, ask on the Crowdin project page or in
+a [discussion](https://github.com/netbirdio/netbird/discussions).
+
 ## Other project repositories
 
 NetBird project is composed of 3 main repositories:

client/android/client.go

@@ -15,6 +15,7 @@ import (
 
 	log "github.com/sirupsen/logrus"
 
+	nbAnonymize "github.com/netbirdio/netbird/client/anonymize"
 	"github.com/netbirdio/netbird/client/iface/device"
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/debug"
@@ -32,6 +33,13 @@ import (
 	types "github.com/netbirdio/netbird/upload-server/types"
 )
 
+// AnonymizeLevelDefault and AnonymizeLevelStrict are the accepted
+// anonymizeLevel values for DebugBundle.
+const (
+	AnonymizeLevelDefault = nbAnonymize.LevelDefaultString
+	AnonymizeLevelStrict  = nbAnonymize.LevelStrictString
+)
+
 // ConnectionListener export internal Listener for mobile
 type ConnectionListener interface {
 	peer.Listener
@@ -278,8 +286,10 @@ func (c *Client) GetTunSettings() (*TunSettings, error) {
 }
 
 // DebugBundle generates a debug bundle, uploads it, and returns the upload key.
-// It works both with and without a running engine.
-func (c *Client) DebugBundle(platformFiles PlatformFiles, anonymize bool) (string, error) {
+// It works both with and without a running engine. anonymizeLevel is "default"
+// or "strict"; strict also anonymizes internal IP ranges, peer names, and
+// WireGuard public keys, and implies anonymize.
+func (c *Client) DebugBundle(platformFiles PlatformFiles, anonymize bool, anonymizeLevel string) (string, error) {
 	cfg, cacheDir, cc := c.stateSnapshot()
 
 	// If the engine hasn't been started, load config from disk
@@ -298,6 +308,7 @@ func (c *Client) DebugBundle(platformFiles PlatformFiles, anonymize bool) (strin
 		InternalConfig: cfg,
 		StatusRecorder: c.recorder,
 		TempDir:        cacheDir,
+		StatePath:      platformFiles.StateFilePath(),
 	}
 
 	if cc != nil {
@@ -321,6 +332,7 @@ func (c *Client) DebugBundle(platformFiles PlatformFiles, anonymize bool) (strin
 		deps,
 		debug.BundleConfig{
 			Anonymize:         anonymize,
+			AnonymizeLevel:    nbAnonymize.ParseLevel(anonymizeLevel),
 			IncludeSystemInfo: true,
 		},
 	)

client/anonymize/anonymize.go

@@ -2,6 +2,7 @@ package anonymize
 
 import (
 	"crypto/rand"
+	"encoding/base64"
 	"fmt"
 	"math/big"
 	"net"
@@ -15,13 +16,88 @@ import (
 
 const anonTLD = ".domain"
 
+// Level selects how much the anonymizer redacts. Levels are ordered: a higher
+// level redacts strictly more. On the wire (protos, flags) levels travel as
+// their string form.
+type Level int
+
+const (
+	// LevelDefault anonymizes public IP addresses, IPv6 ULA, domains, and MAC
+	// addresses. Internal IPv4 ranges (RFC 1918, CGNAT, link-local) are
+	// preserved so support can reason about the real topology.
+	LevelDefault Level = iota
+	// LevelStrict additionally anonymizes internal IP ranges, peer names, and
+	// WireGuard public keys.
+	LevelStrict
+)
+
+// LevelDefaultString and LevelStrictString are the wire forms of the levels,
+// for boundaries that pass levels as strings (flags, protos, mobile bindings).
+const (
+	LevelDefaultString = "default"
+	LevelStrictString  = "strict"
+)
+
+// ParseLevel maps s to a Level. Empty means LevelDefault; anything
+// unrecognized maps to LevelStrict so an unknown request never yields less
+// anonymization than intended.
+func ParseLevel(s string) Level {
+	switch strings.ToLower(s) {
+	case "", LevelDefaultString:
+		return LevelDefault
+	default:
+		return LevelStrict
+	}
+}
+
+// String returns the wire form of the level: "default" or "strict".
+func (l Level) String() string {
+	if l >= LevelStrict {
+		return LevelStrictString
+	}
+	return LevelDefaultString
+}
+
+// protectedDomains are NetBird-operated suffixes that stay recognizable in an
+// anonymized bundle. At LevelStrict the labels in front of them (the peer
+// name) are still replaced, except under netbird.io, which only hosts
+// NetBird infrastructure (api, signal, flow), never peer names.
+var protectedDomains = []string{"netbird.io", "netbird.selfhosted", "netbird.cloud", "netbird.stage"}
+
+const infraDomain = "netbird.io"
+
+var (
+	macColonRegex = regexp.MustCompile(`\b[0-9a-fA-F]{2}(?::[0-9a-fA-F]{2}){5}\b`)
+	macDashRegex  = regexp.MustCompile(`\b[0-9a-fA-F]{2}(?:-[0-9a-fA-F]{2}){5}\b`)
+	wgKeyRegex    = regexp.MustCompile(`\b[A-Za-z0-9+/]{43}=`)
+)
+
 type Anonymizer struct {
 	ipAnonymizer     map[netip.Addr]netip.Addr
 	domainAnonymizer map[string]string
-	currentAnonIPv4  netip.Addr
-	currentAnonIPv6  netip.Addr
-	startAnonIPv4    netip.Addr
-	startAnonIPv6    netip.Addr
+	// domainOrder caches the keys of domainAnonymizer sorted longest-first
+	// for AnonymizeString; it is rebuilt when the map gains entries.
+	domainOrder     []string
+	labelAnonymizer map[string]string
+	labelAnonymized map[string]struct{}
+	labelCounter    uint32
+	macAnonymizer   map[string]string
+	macCounter      uint32
+	wgKeyAnonymizer map[string]string
+	wgKeyAnonymized map[string]struct{}
+	currentAnonIPv4 netip.Addr
+	currentAnonIPv6 netip.Addr
+	startAnonIPv4   netip.Addr
+	startAnonIPv6   netip.Addr
+
+	// LevelStrict also anonymizes internal ranges (RFC 1918, CGNAT,
+	// link-local), replacing them from the dedicated internal pools below so
+	// a reader can still tell an internal address from a public one.
+	level                   Level
+	currentAnonInternalIPv4 netip.Addr
+	currentAnonInternalIPv6 netip.Addr
+	startAnonInternalIPv4   netip.Addr
+	startAnonInternalIPv6   netip.Addr
 
 	domainKeyRegex *regexp.Regexp
 }
@@ -32,25 +108,50 @@ func DefaultAddresses() (netip.Addr, netip.Addr) {
 	return netip.AddrFrom4([4]byte{198, 51, 100, 0}), netip.MustParseAddr("2001:db8:ffff::")
 }
 
+// InternalAddresses returns the pool starts used in strict mode for internal
+// ranges. Both are reserved ranges that cannot collide with real addressing:
+// 198.18.0.0 (RFC 2544 benchmarking), 2001:db8:1:: (RFC 3849 documentation).
+func InternalAddresses() (netip.Addr, netip.Addr) {
+	return netip.AddrFrom4([4]byte{198, 18, 0, 0}), netip.MustParseAddr("2001:db8:1::")
+}
+
 func NewAnonymizer(startIPv4, startIPv6 netip.Addr) *Anonymizer {
+	internalIPv4, internalIPv6 := InternalAddresses()
 	return &Anonymizer{
 		ipAnonymizer:     map[netip.Addr]netip.Addr{},
 		domainAnonymizer: map[string]string{},
+		labelAnonymizer:  map[string]string{},
+		labelAnonymized:  map[string]struct{}{},
+		macAnonymizer:    map[string]string{},
+		wgKeyAnonymizer:  map[string]string{},
+		wgKeyAnonymized:  map[string]struct{}{},
 		currentAnonIPv4:  startIPv4,
 		currentAnonIPv6:  startIPv6,
 		startAnonIPv4:    startIPv4,
 		startAnonIPv6:    startIPv6,
 
+		level:                   LevelDefault,
+		currentAnonInternalIPv4: internalIPv4,
+		currentAnonInternalIPv6: internalIPv6,
+		startAnonInternalIPv4:   internalIPv4,
+		startAnonInternalIPv6:   internalIPv6,
+
 		domainKeyRegex: regexp.MustCompile(`\bdomain=([^\s,:"]+)`),
 	}
 }
 
+// SetLevel selects the anonymization level. The zero value of a new
+// Anonymizer is LevelDefault.
+func (a *Anonymizer) SetLevel(level Level) {
+	a.level = level
+}
+
 func (a *Anonymizer) AnonymizeIP(ip netip.Addr) netip.Addr {
+	// Normalize 4-in-6 addresses so ::ffff:192.168.1.1 classifies and maps
+	// like 192.168.1.1.
+	ip = ip.Unmap()
+
 	if ip.IsLoopback() ||
-		ip.IsLinkLocalUnicast() ||
-		ip.IsLinkLocalMulticast() ||
-		ip.IsInterfaceLocalMulticast() ||
-		(ip.Is4() && ip.IsPrivate()) ||
 		ip.IsUnspecified() ||
 		ip.IsMulticast() ||
 		isWellKnown(ip) ||
@@ -59,18 +160,100 @@ func (a *Anonymizer) AnonymizeIP(ip netip.Addr) netip.Addr {
 		return ip
 	}
 
+	if isInternal(ip) && a.level < LevelStrict {
+		return ip
+	}
+
 	if _, ok := a.ipAnonymizer[ip]; !ok {
-		if ip.Is4() {
-			a.ipAnonymizer[ip] = a.currentAnonIPv4
-			a.currentAnonIPv4 = a.currentAnonIPv4.Next()
-		} else {
-			a.ipAnonymizer[ip] = a.currentAnonIPv6
-			a.currentAnonIPv6 = a.currentAnonIPv6.Next()
-		}
+		a.ipAnonymizer[ip] = a.nextAnonIP(ip)
 	}
 	return a.ipAnonymizer[ip]
 }
 
+func (a *Anonymizer) nextAnonIP(ip netip.Addr) netip.Addr {
+	// At the strict level, internal addresses (including IPv6 ULA, matched
+	// by IsPrivate) come from the internal pools so they remain recognizable
+	// as internal without disclosing the real values.
+	if a.level >= LevelStrict && (isInternal(ip) || ip.IsPrivate()) {
+		if ip.Is4() {
+			anon := a.currentAnonInternalIPv4
+			a.currentAnonInternalIPv4 = a.currentAnonInternalIPv4.Next()
+			return anon
+		}
+		anon := a.currentAnonInternalIPv6
+		a.currentAnonInternalIPv6 = a.currentAnonInternalIPv6.Next()
+		return anon
+	}
+
+	if ip.Is4() {
+		anon := a.currentAnonIPv4
+		a.currentAnonIPv4 = a.currentAnonIPv4.Next()
+		return anon
+	}
+	anon := a.currentAnonIPv6
+	a.currentAnonIPv6 = a.currentAnonIPv6.Next()
+	return anon
+}
+
+// AnonymizeMAC replaces a MAC address with a consistent placeholder from the
+// locally administered range starting at 02:00:00:00:00:01, at every
+// anonymization level. Broadcast, multicast, all-zero, and already assigned
+// placeholder addresses are preserved. The colon and dash spellings of the
+// same address share one placeholder; the output keeps the input's separator.
+func (a *Anonymizer) AnonymizeMAC(mac string) string {
+	hw, err := net.ParseMAC(mac)
+	if err != nil || len(hw) != 6 {
+		return mac
+	}
+
+	if isWellKnownMAC(hw) || a.isAnonymizedMAC(hw) {
+		return mac
+	}
+
+	key := hw.String()
+	anon, ok := a.macAnonymizer[key]
+	if !ok {
+		a.macCounter++
+		anon = fmt.Sprintf("02:00:00:%02x:%02x:%02x", byte(a.macCounter>>16), byte(a.macCounter>>8), byte(a.macCounter))
+		a.macAnonymizer[key] = anon
+	}
+
+	if strings.Contains(mac, "-") {
+		anon = strings.ReplaceAll(anon, ":", "-")
+	}
+	return anon
+}
+
+// isAnonymizedMAC reports whether hw is a placeholder this anonymizer already
+// handed out, so a second pass over anonymized output leaves it unchanged.
+func (a *Anonymizer) isAnonymizedMAC(hw net.HardwareAddr) bool {
+	if hw[0] != 0x02 || hw[1] != 0 || hw[2] != 0 {
+		return false
+	}
+	value := uint32(hw[3])<<16 | uint32(hw[4])<<8 | uint32(hw[5])
+	return value <= a.macCounter
+}
+
+// AnonymizeWGKey replaces a WireGuard public key with a consistent random
+// placeholder of the same shape. Keys are only anonymized at LevelStrict;
+// placeholders already handed out pass through unchanged.
+func (a *Anonymizer) AnonymizeWGKey(key string) string {
+	if a.level < LevelStrict || !looksLikeWGKey(key) {
+		return key
+	}
+	if _, ok := a.wgKeyAnonymized[key]; ok {
+		return key
+	}
+
+	anon, ok := a.wgKeyAnonymizer[key]
+	if !ok {
+		anon = generateAnonymousKey()
+		a.wgKeyAnonymizer[key] = anon
+		a.wgKeyAnonymized[anon] = struct{}{}
+	}
+	return anon
+}
+
 func (a *Anonymizer) AnonymizeUDPAddr(addr net.UDPAddr) net.UDPAddr {
 	// Convert IP to netip.Addr
 	ip, ok := netip.AddrFromSlice(addr.IP)
@@ -89,12 +272,12 @@ func (a *Anonymizer) AnonymizeUDPAddr(addr net.UDPAddr) net.UDPAddr {
 
 // isInAnonymizedRange checks if an IP is within the range of already assigned anonymized IPs
 func (a *Anonymizer) isInAnonymizedRange(ip netip.Addr) bool {
-	if ip.Is4() && ip.Compare(a.startAnonIPv4) >= 0 && ip.Compare(a.currentAnonIPv4) <= 0 {
-		return true
-	} else if !ip.Is4() && ip.Compare(a.startAnonIPv6) >= 0 && ip.Compare(a.currentAnonIPv6) <= 0 {
-		return true
+	if ip.Is4() {
+		return inPoolRange(ip, a.startAnonIPv4, a.currentAnonIPv4) ||
+			inPoolRange(ip, a.startAnonInternalIPv4, a.currentAnonInternalIPv4)
 	}
-	return false
+	return inPoolRange(ip, a.startAnonIPv6, a.currentAnonIPv6) ||
+		inPoolRange(ip, a.startAnonInternalIPv6, a.currentAnonInternalIPv6)
 }
 
 func (a *Anonymizer) AnonymizeIPString(ip string) string {
@@ -118,14 +301,17 @@ func (a *Anonymizer) AnonymizeDomain(domain string) string {
 		baseDomain = domain[:len(domain)-1]
 	}
 
-	if strings.HasSuffix(baseDomain, "netbird.io") ||
-		strings.HasSuffix(baseDomain, "netbird.selfhosted") ||
-		strings.HasSuffix(baseDomain, "netbird.cloud") ||
-		strings.HasSuffix(baseDomain, "netbird.stage") ||
-		strings.HasSuffix(baseDomain, anonTLD) {
+	if strings.HasSuffix(baseDomain, anonTLD) {
 		return domain
 	}
 
+	if suffix := protectedSuffix(baseDomain); suffix != "" {
+		if a.level < LevelStrict || baseDomain == suffix || suffix == infraDomain {
+			return domain
+		}
+		return withTrailingDot(a.anonymizePeerName(baseDomain, suffix), hasDot)
+	}
+
 	parts := strings.Split(baseDomain, ".")
 	if len(parts) < 2 {
 		return domain
@@ -141,12 +327,53 @@ func (a *Anonymizer) AnonymizeDomain(domain string) string {
 	}
 
 	result := strings.Replace(baseDomain, baseForLookup, anonymized, 1)
-	if hasDot {
-		result += "."
+	if a.level >= LevelStrict && len(parts) > 2 {
+		prefix := strings.TrimSuffix(baseDomain, "."+baseForLookup)
+		result = a.anonymizeLabels(prefix, "host") + "." + anonymized
+		// The full mapping feeds AnonymizeString so seeded FQDNs are caught
+		// in log lines as a whole, labels included.
+		a.domainAnonymizer[baseDomain] = result
+	}
+	return withTrailingDot(result, hasDot)
+}
+
+// anonymizePeerName replaces the labels in front of a protected suffix with
+// numbered peer placeholders, keeping the suffix, and records the full
+// mapping for string replacement in logs. The numbering keeps a peer
+// recognizable across the whole bundle without disclosing its name.
+func (a *Anonymizer) anonymizePeerName(baseDomain, suffix string) string {
+	prefix := strings.TrimSuffix(baseDomain, "."+suffix)
+	result := a.anonymizeLabels(prefix, "peer") + "." + suffix
+	if result != baseDomain {
+		a.domainAnonymizer[baseDomain] = result
 	}
 	return result
 }
 
+// anonymizeLabels replaces each dot-separated label with a consistent
+// numbered placeholder ("<placeholder>-<n>"). Wildcard labels and
+// placeholders already handed out pass through unchanged.
+func (a *Anonymizer) anonymizeLabels(prefix, placeholder string) string {
+	labels := strings.Split(prefix, ".")
+	for i, label := range labels {
+		if label == "*" {
+			continue
+		}
+		if _, ok := a.labelAnonymized[label]; ok {
+			continue
+		}
+		anon, ok := a.labelAnonymizer[label]
+		if !ok {
+			a.labelCounter++
+			anon = fmt.Sprintf("%s-%d", placeholder, a.labelCounter)
+			a.labelAnonymizer[label] = anon
+			a.labelAnonymized[anon] = struct{}{}
+		}
+		labels[i] = anon
+	}
+	return strings.Join(labels, ".")
+}
+
 func (a *Anonymizer) AnonymizeURI(uri string) string {
 	u, err := url.Parse(uri)
 	if err != nil {
@@ -181,16 +408,70 @@ func (a *Anonymizer) AnonymizeString(str string) string {
 	str = ipv4Regex.ReplaceAllStringFunc(str, a.AnonymizeIPString)
 	str = ipv6Regex.ReplaceAllStringFunc(str, a.AnonymizeIPString)
 
-	for domain, anonDomain := range a.domainAnonymizer {
-		str = strings.ReplaceAll(str, domain, anonDomain)
+	for _, domain := range a.sortedDomains() {
+		str = strings.ReplaceAll(str, domain, a.domainAnonymizer[domain])
 	}
 
 	str = a.AnonymizeSchemeURI(str)
 	str = a.AnonymizeDNSLogLine(str)
 
+	// MAC handling runs after the IP passes so preserved IPv6 addresses are
+	// already out of the way; the separator guard skips matches embedded in a
+	// longer colon- or dash-separated sequence (such as an IPv6 tail).
+	str = a.anonymizeMACsInString(str, macColonRegex, ':')
+	str = a.anonymizeMACsInString(str, macDashRegex, '-')
+
+	if a.level >= LevelStrict {
+		str = wgKeyRegex.ReplaceAllStringFunc(str, a.AnonymizeWGKey)
+	}
+
 	return str
 }
 
+// sortedDomains returns the domain mappings longest-first, so a full-FQDN
+// mapping (strict level) is applied before the base-domain mapping it
+// contains. The order is rebuilt only when domainAnonymizer has grown.
+func (a *Anonymizer) sortedDomains() []string {
+	if len(a.domainOrder) == len(a.domainAnonymizer) {
+		return a.domainOrder
+	}
+
+	a.domainOrder = a.domainOrder[:0]
+	for domain := range a.domainAnonymizer {
+		a.domainOrder = append(a.domainOrder, domain)
+	}
+	slices.SortFunc(a.domainOrder, func(x, y string) int {
+		if d := len(y) - len(x); d != 0 {
+			return d
+		}
+		return strings.Compare(x, y)
+	})
+	return a.domainOrder
+}
+
+// anonymizeMACsInString replaces MAC addresses matched by re, skipping
+// matches that directly adjoin another sep so a six-group run inside a longer
+// separated sequence is left alone.
+func (a *Anonymizer) anonymizeMACsInString(str string, re *regexp.Regexp, sep byte) string {
+	matches := re.FindAllStringIndex(str, -1)
+	if len(matches) == 0 {
+		return str
+	}
+
+	var b strings.Builder
+	last := 0
+	for _, m := range matches {
+		if (m[0] > 0 && str[m[0]-1] == sep) || (m[1] < len(str) && str[m[1]] == sep) {
+			continue
+		}
+		b.WriteString(str[last:m[0]])
+		b.WriteString(a.AnonymizeMAC(str[m[0]:m[1]]))
+		last = m[1]
+	}
+	b.WriteString(str[last:])
+	return b.String()
+}
+
 // AnonymizeSchemeURI finds and anonymizes URIs with ws, wss, rel, rels, stun, stuns, turn, and turns schemes.
 func (a *Anonymizer) AnonymizeSchemeURI(text string) string {
 	re := regexp.MustCompile(`(?i)\b(wss?://|rels?://|stuns?:|turns?:|https?://)\S+\b`)
@@ -239,10 +520,79 @@ func isWellKnown(addr netip.Addr) bool {
 		"128.0.0.0", "8000::", // 2nd split subnet for default routes
 	}
 
-	if slices.Contains(wellKnown, addr.String()) {
+	return slices.Contains(wellKnown, addr.String())
+}
+
+// isInternal reports whether ip identifies a host only within the local
+// network: IPv4 private (RFC 1918), CGNAT (RFC 6598), and link-local (v4 and
+// v6). These are preserved at the default level so support can reason about
+// the real topology, and replaced from the internal pools at the strict
+// level. IPv6 ULA is deliberately not internal: its random global ID uniquely
+// fingerprints the network, so it is anonymized at every level.
+func isInternal(ip netip.Addr) bool {
+	return (ip.Is4() && ip.IsPrivate()) ||
+		ip.IsLinkLocalUnicast() ||
+		isCGNAT(ip)
+}
+
+func inPoolRange(ip, start, current netip.Addr) bool {
+	return ip.Compare(start) >= 0 && ip.Compare(current) <= 0
+}
+
+// isWellKnownMAC reports whether hw carries no stable host identity: all-zero
+// or a group address (broadcast and multicast).
+func isWellKnownMAC(hw net.HardwareAddr) bool {
+	if hw[0]&1 == 1 {
 		return true
 	}
+	for _, b := range hw {
+		if b != 0 {
+			return false
+		}
+	}
+	return true
+}
 
+// looksLikeWGKey reports whether s has the shape of a WireGuard key:
+// 44 base64 characters decoding to 32 bytes.
+func looksLikeWGKey(s string) bool {
+	if len(s) != 44 || s[43] != '=' {
+		return false
+	}
+	decoded, err := base64.StdEncoding.DecodeString(s)
+	return err == nil && len(decoded) == 32
+}
+
+func generateAnonymousKey() string {
+	buf := make([]byte, 32)
+	if _, err := rand.Read(buf); err != nil {
+		return strings.Repeat("A", 43) + "="
+	}
+	return base64.StdEncoding.EncodeToString(buf)
+}
+
+// protectedSuffix returns the protected NetBird suffix baseDomain ends with,
+// or empty. The match is label-anchored so an unrelated domain that merely
+// ends in the same characters is not preserved.
+func protectedSuffix(baseDomain string) string {
+	for _, d := range protectedDomains {
+		if baseDomain == d || strings.HasSuffix(baseDomain, "."+d) {
+			return d
+		}
+	}
+	return ""
+}
+
+func withTrailingDot(domain string, hasDot bool) string {
+	if hasDot {
+		return domain + "."
+	}
+	return domain
+}
+
+// isCGNAT reports whether addr is in 100.64.0.0/10 (RFC 6598), the range
+// NetBird assigns overlay peer addresses from.
+func isCGNAT(addr netip.Addr) bool {
 	cgnatRangeStart := netip.AddrFrom4([4]byte{100, 64, 0, 0})
 	cgnatRange := netip.PrefixFrom(cgnatRangeStart, 10)
 

client/anonymize/anonymize_test.go

@@ -1,8 +1,11 @@
 package anonymize_test
 
 import (
+	"bytes"
+	"encoding/base64"
 	"net/netip"
 	"regexp"
+	"strings"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -44,6 +47,301 @@ func TestAnonymizeIP(t *testing.T) {
 	}
 }
 
+func TestParseLevel(t *testing.T) {
+	tests := []struct {
+		input  string
+		expect anonymize.Level
+	}{
+		{"", anonymize.LevelDefault},
+		{"default", anonymize.LevelDefault},
+		{"DEFAULT", anonymize.LevelDefault},
+		{"strict", anonymize.LevelStrict},
+		{"STRICT", anonymize.LevelStrict},
+		// Unknown values must never yield less anonymization than requested.
+		{"garbage", anonymize.LevelStrict},
+	}
+
+	for _, tc := range tests {
+		t.Run("input="+tc.input, func(t *testing.T) {
+			assert.Equal(t, tc.expect, anonymize.ParseLevel(tc.input), "parsed level should match")
+		})
+	}
+}
+
+func TestAnonymizeIP_DefaultLevelInternalRanges(t *testing.T) {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+
+	tests := []struct {
+		name   string
+		ip     string
+		expect string
+	}{
+		{"RFC1918 10/8", "10.1.2.3", "10.1.2.3"},
+		{"RFC1918 172.16/12", "172.16.5.5", "172.16.5.5"},
+		{"RFC1918 192.168/16", "192.168.1.1", "192.168.1.1"},
+		{"CGNAT", "100.64.0.5", "100.64.0.5"},
+		{"IPv4 link-local", "169.254.1.1", "169.254.1.1"},
+		{"IPv6 link-local", "fe80::1", "fe80::1"},
+		// ULA is anonymized even at the default level: its random global ID
+		// uniquely fingerprints the network, unlike shared RFC 1918 space.
+		{"IPv6 ULA", "fd12:3456:789a::1", "2001:db8:ffff::"},
+		// 4-in-6 addresses classify like their unmapped IPv4 form.
+		{"4-in-6 RFC1918", "::ffff:192.168.1.1", "192.168.1.1"},
+		{"4-in-6 CGNAT", "::ffff:100.64.0.5", "100.64.0.5"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			result := anonymizer.AnonymizeIP(netip.MustParseAddr(tc.ip))
+			assert.Equal(t, tc.expect, result.String(), "default level should preserve internal ranges except ULA")
+		})
+	}
+}
+
+func TestAnonymizeIP_StrictLevel(t *testing.T) {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+	anonymizer.SetLevel(anonymize.LevelStrict)
+
+	// Order matters: internal pool addresses are assigned sequentially.
+	tests := []struct {
+		name   string
+		ip     string
+		expect string
+	}{
+		{"RFC1918 192.168/16", "192.168.1.1", "198.18.0.0"},
+		{"Second RFC1918", "192.168.1.2", "198.18.0.1"},
+		{"Repeated RFC1918", "192.168.1.1", "198.18.0.0"},
+		{"RFC1918 10/8", "10.1.2.3", "198.18.0.2"},
+		{"RFC1918 172.16/12", "172.16.5.5", "198.18.0.3"},
+		{"CGNAT", "100.64.0.5", "198.18.0.4"},
+		{"IPv4 link-local", "169.254.1.1", "198.18.0.5"},
+		{"Public IPv4 uses public pool", "1.2.3.4", "198.51.100.0"},
+		{"IPv6 link-local", "fe80::1", "2001:db8:1::"},
+		{"IPv6 ULA", "fd12:3456:789a::1", "2001:db8:1::1"},
+		{"Public IPv6 uses public pool", "2607:f8b0:4005:805::200e", "2001:db8:ffff::"},
+		{"Loopback IPv4", "127.0.0.1", "127.0.0.1"},
+		{"Loopback IPv6", "::1", "::1"},
+		{"Unspecified", "0.0.0.0", "0.0.0.0"},
+		{"Multicast", "224.0.0.251", "224.0.0.251"},
+		{"Well known resolver", "8.8.8.8", "8.8.8.8"},
+		{"Well known split marker", "128.0.0.0", "128.0.0.0"},
+		{"In internal pool range", "198.18.0.3", "198.18.0.3"},
+		{"In public pool range", "198.51.100.0", "198.51.100.0"},
+		{"4-in-6 repeated RFC1918", "::ffff:192.168.1.1", "198.18.0.0"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			result := anonymizer.AnonymizeIP(netip.MustParseAddr(tc.ip))
+			assert.Equal(t, tc.expect, result.String(), "strict level should replace internal ranges from the internal pools")
+		})
+	}
+}
+
+func TestAnonymizeString_StrictInternalIPs(t *testing.T) {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+	anonymizer.SetLevel(anonymize.LevelStrict)
+
+	input := "route 10.20.30.0/24 via 192.168.1.1 dev eth0 src 100.64.0.7"
+	firstPass := anonymizer.AnonymizeString(input)
+	secondPass := anonymizer.AnonymizeString(firstPass)
+
+	assert.NotContains(t, firstPass, "10.20.30.0", "private network address should be anonymized")
+	assert.NotContains(t, firstPass, "192.168.1.1", "private gateway should be anonymized")
+	assert.NotContains(t, firstPass, "100.64.0.7", "CGNAT address should be anonymized")
+	assert.Contains(t, firstPass, "/24", "prefix length should be preserved")
+	assert.Equal(t, firstPass, secondPass, "second pass should not further anonymize the string")
+}
+
+func TestAnonymizeMAC(t *testing.T) {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+
+	first := anonymizer.AnonymizeMAC("aa:bb:cc:dd:ee:0f")
+	assert.Equal(t, "02:00:00:00:00:01", first, "first MAC should get the first placeholder")
+	assert.Equal(t, first, anonymizer.AnonymizeMAC("aa:bb:cc:dd:ee:0f"), "repeated MAC should map to the same placeholder")
+	assert.Equal(t, first, anonymizer.AnonymizeMAC("AA:BB:CC:DD:EE:0F"), "case should not affect the mapping")
+	assert.Equal(t, "02-00-00-00-00-01", anonymizer.AnonymizeMAC("AA-BB-CC-DD-EE-0F"), "dash form should keep its separator but share the mapping")
+
+	second := anonymizer.AnonymizeMAC("10:22:33:44:55:66")
+	assert.Equal(t, "02:00:00:00:00:02", second, "second distinct MAC should get the next placeholder")
+
+	tests := []struct {
+		name string
+		mac  string
+	}{
+		{"Broadcast", "ff:ff:ff:ff:ff:ff"},
+		{"IPv4 multicast", "01:00:5e:00:00:fb"},
+		{"IPv6 multicast", "33:33:00:00:00:01"},
+		{"All zero", "00:00:00:00:00:00"},
+		{"Assigned placeholder", "02:00:00:00:00:01"},
+		{"Invalid", "not-a-mac"},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			assert.Equal(t, tc.mac, anonymizer.AnonymizeMAC(tc.mac), "should be preserved")
+		})
+	}
+}
+
+func TestAnonymizeString_MACAddresses(t *testing.T) {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+
+	tests := []struct {
+		name   string
+		input  string
+		expect string
+	}{
+		{
+			name:   "nftables ether rule",
+			input:  "ether saddr aa:bb:cc:dd:ee:ff drop",
+			expect: "ether saddr 02:00:00:00:00:01 drop",
+		},
+		{
+			name:   "Windows dash form",
+			input:  "Physical Address : AA-BB-CC-DD-EE-FF",
+			expect: "Physical Address : 02-00-00-00-00-01",
+		},
+		{
+			name:   "IPv6 address tail is not treated as MAC",
+			input:  "addr fe80:0:11:22:33:44:55:66 scope link",
+			expect: "addr fe80:0:11:22:33:44:55:66 scope link",
+		},
+		{
+			name:   "broadcast MAC preserved",
+			input:  "dst ff:ff:ff:ff:ff:ff type ARP",
+			expect: "dst ff:ff:ff:ff:ff:ff type ARP",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			result := anonymizer.AnonymizeString(tc.input)
+			assert.Equal(t, tc.expect, result, "MAC addresses should be anonymized at every level")
+			assert.Equal(t, result, anonymizer.AnonymizeString(result), "second pass should not change the result")
+		})
+	}
+}
+
+func TestAnonymizeWGKey(t *testing.T) {
+	key := base64.StdEncoding.EncodeToString(bytes.Repeat([]byte{0x42}, 32))
+
+	t.Run("default level preserves keys", func(t *testing.T) {
+		anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+		assert.Equal(t, key, anonymizer.AnonymizeWGKey(key), "default level should not touch WireGuard keys")
+	})
+
+	t.Run("strict level replaces keys", func(t *testing.T) {
+		anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+		anonymizer.SetLevel(anonymize.LevelStrict)
+
+		anon := anonymizer.AnonymizeWGKey(key)
+		assert.NotEqual(t, key, anon, "strict level should replace the key")
+		assert.Regexp(t, `^[A-Za-z0-9+/]{43}=$`, anon, "placeholder should keep the WireGuard key shape")
+		assert.Equal(t, anon, anonymizer.AnonymizeWGKey(key), "repeated key should map to the same placeholder")
+		assert.Equal(t, anon, anonymizer.AnonymizeWGKey(anon), "an assigned placeholder should pass through unchanged")
+
+		assert.Equal(t, "not-a-key", anonymizer.AnonymizeWGKey("not-a-key"), "non-key values should be preserved")
+	})
+}
+
+func TestAnonymizeString_WGKeys(t *testing.T) {
+	key := base64.StdEncoding.EncodeToString(bytes.Repeat([]byte{0x42}, 32))
+	input := "peer " + key + " handshake completed"
+
+	t.Run("default level preserves keys", func(t *testing.T) {
+		anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+		assert.Equal(t, input, anonymizer.AnonymizeString(input), "default level should not touch WireGuard keys in strings")
+	})
+
+	t.Run("strict level replaces keys", func(t *testing.T) {
+		anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+		anonymizer.SetLevel(anonymize.LevelStrict)
+
+		firstPass := anonymizer.AnonymizeString(input)
+		assert.NotContains(t, firstPass, key, "the key should not survive strict anonymization")
+		assert.Equal(t, anonymizer.AnonymizeWGKey(key), extractKey(t, firstPass), "string replacement should be consistent with AnonymizeWGKey")
+		assert.Equal(t, firstPass, anonymizer.AnonymizeString(firstPass), "second pass should not change the result")
+	})
+}
+
+func extractKey(t *testing.T, logLine string) string {
+	t.Helper()
+	fields := strings.Fields(logLine)
+	require.Len(t, fields, 4, "log line should keep its structure")
+	return fields[1]
+}
+
+func TestAnonymizeDomain_StrictLevel(t *testing.T) {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+	anonymizer.SetLevel(anonymize.LevelStrict)
+
+	t.Run("netbird peer name", func(t *testing.T) {
+		result := anonymizer.AnonymizeDomain("my-laptop.netbird.cloud")
+		assert.Regexp(t, `^peer-\d+\.netbird\.cloud$`, result, "peer name should be anonymized, suffix kept")
+		assert.NotContains(t, result, "my-laptop", "the peer name should not survive")
+		assert.Equal(t, result, anonymizer.AnonymizeDomain("my-laptop.netbird.cloud"), "repeated domain should map consistently")
+		assert.Equal(t, result, anonymizer.AnonymizeDomain(result), "an anonymized domain should pass through unchanged")
+	})
+
+	t.Run("bare netbird domain", func(t *testing.T) {
+		assert.Equal(t, "netbird.cloud", anonymizer.AnonymizeDomain("netbird.cloud"), "the bare protected suffix should be preserved")
+	})
+
+	t.Run("netbird infrastructure preserved", func(t *testing.T) {
+		assert.Equal(t, "api.netbird.io", anonymizer.AnonymizeDomain("api.netbird.io"),
+			"netbird.io hosts infrastructure, not peer names, and should stay readable")
+	})
+
+	t.Run("leading labels of other domains", func(t *testing.T) {
+		result := anonymizer.AnonymizeDomain("host1.corp.example.com")
+		assert.Regexp(t, `^host-\d+\.host-\d+\.anon-[a-zA-Z0-9]+\.domain$`, result, "every label should be anonymized")
+		for _, label := range []string{"host1", "corp", "example"} {
+			assert.NotContains(t, result, label, "no original label should survive")
+		}
+		assert.Equal(t, result, anonymizer.AnonymizeDomain("host1.corp.example.com"), "repeated domain should map consistently")
+	})
+
+	t.Run("same label maps consistently across domains", func(t *testing.T) {
+		first := anonymizer.AnonymizeDomain("shared.one.com")
+		second := anonymizer.AnonymizeDomain("shared.two.com")
+		assert.Equal(t, strings.Split(first, ".")[0], strings.Split(second, ".")[0], "the shared host label should get one placeholder")
+	})
+
+	t.Run("wildcard label preserved", func(t *testing.T) {
+		result := anonymizer.AnonymizeDomain("*.example.com")
+		assert.Regexp(t, `^\*\.anon-[a-zA-Z0-9]+\.domain$`, result, "the wildcard label should stay a wildcard")
+	})
+}
+
+func TestAnonymizeDomain_DefaultLevelKeepsPeerNames(t *testing.T) {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+
+	assert.Equal(t, "my-laptop.netbird.cloud", anonymizer.AnonymizeDomain("my-laptop.netbird.cloud"),
+		"default level should preserve netbird FQDNs including the peer name")
+	assert.Regexp(t, `^sub\.anon-[a-zA-Z0-9]+\.domain$`, anonymizer.AnonymizeDomain("sub.example.com"),
+		"default level should keep subdomain labels")
+}
+
+func TestAnonymizeString_StrictPeerNames(t *testing.T) {
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+	anonymizer.SetLevel(anonymize.LevelStrict)
+
+	// Seed like the bundle generator does from the status: base first, then
+	// the full FQDN, so replacement must prefer the longer mapping.
+	anonBase := anonymizer.AnonymizeDomain("example.com")
+	anonPeer := anonymizer.AnonymizeDomain("peer1.netbird.cloud")
+	anonHost := anonymizer.AnonymizeDomain("host1.example.com")
+
+	logLine := "connected to peer1.netbird.cloud via host1.example.com endpoint"
+	firstPass := anonymizer.AnonymizeString(logLine)
+	assert.NotContains(t, firstPass, "peer1", "the peer name should not survive in logs")
+	assert.NotContains(t, firstPass, "host1", "the host label should not survive in logs")
+	assert.Contains(t, firstPass, anonPeer, "the seeded peer mapping should be applied")
+	assert.Contains(t, firstPass, anonHost, "the seeded host mapping should be applied, not just the base mapping")
+	assert.NotContains(t, firstPass, "host1."+anonBase, "the base mapping must not preempt the longer FQDN mapping")
+	assert.Equal(t, firstPass, anonymizer.AnonymizeString(firstPass), "second pass should not change the result")
+}
+
 func TestAnonymizeDNSLogLine(t *testing.T) {
 	anonymizer := anonymize.NewAnonymizer(netip.Addr{}, netip.Addr{})
 	tests := []struct {

client/cmd/debug.go

@@ -27,8 +27,8 @@ import (
 const errCloseConnection = "Failed to close connection: %v"
 
 var (
-	logFileCount        uint32
-	systemInfoFlag      bool
+	logFileCount             uint32
+	systemInfoFlag           bool
 	uploadBundleFlag         bool
 	uploadBundleURLFlag      string
 	uploadBundleInsecureFlag bool
@@ -156,6 +156,11 @@ func debugConfigDump(cmd *cobra.Command, _ []string) error {
 // request. Returns an error if the RPC fails or if the daemon reports
 // an upload failure reason.
 func debugBundle(cmd *cobra.Command, _ []string) error {
+	anonymizeEnabled, anonymizeLevel, err := effectiveAnonymize()
+	if err != nil {
+		return err
+	}
+
 	conn, err := getClient(cmd)
 	if err != nil {
 		return err
@@ -168,10 +173,11 @@ func debugBundle(cmd *cobra.Command, _ []string) error {
 
 	client := proto.NewDaemonServiceClient(conn)
 	request := &proto.DebugBundleRequest{
-		Anonymize:    anonymizeFlag,
-		SystemInfo:   systemInfoFlag,
-		LogFileCount: logFileCount,
-		CliVersion:   version.NetbirdVersion(),
+		Anonymize:      anonymizeEnabled,
+		AnonymizeLevel: anonymizeLevel.String(),
+		SystemInfo:     systemInfoFlag,
+		LogFileCount:   logFileCount,
+		CliVersion:     version.NetbirdVersion(),
 	}
 	if uploadBundleFlag {
 		request.UploadURL = uploadBundleURLFlag
@@ -229,6 +235,11 @@ func runForDuration(cmd *cobra.Command, args []string) error {
 		return fmt.Errorf("invalid duration format: %v", err)
 	}
 
+	anonymizeEnabled, anonymizeLevel, err := effectiveAnonymize()
+	if err != nil {
+		return err
+	}
+
 	conn, err := getClient(cmd)
 	if err != nil {
 		return err
@@ -368,10 +379,11 @@ func runForDuration(cmd *cobra.Command, args []string) error {
 	cmd.Println("Creating debug bundle...")
 
 	request := &proto.DebugBundleRequest{
-		Anonymize:    anonymizeFlag,
-		SystemInfo:   systemInfoFlag,
-		LogFileCount: logFileCount,
-		CliVersion:   version.NetbirdVersion(),
+		Anonymize:      anonymizeEnabled,
+		AnonymizeLevel: anonymizeLevel.String(),
+		SystemInfo:     systemInfoFlag,
+		LogFileCount:   logFileCount,
+		CliVersion:     version.NetbirdVersion(),
 	}
 	if uploadBundleFlag {
 		request.UploadURL = uploadBundleURLFlag

client/cmd/login.go

@@ -5,7 +5,6 @@ import (
 	"fmt"
 	"os"
 	"os/user"
-	"runtime"
 	"strings"
 
 	log "github.com/sirupsen/logrus"
@@ -121,7 +120,7 @@ func doDaemonLogin(ctx context.Context, cmd *cobra.Command, providedSetupKey str
 	loginRequest := proto.LoginRequest{
 		SetupKey:            providedSetupKey,
 		ManagementUrl:       managementURL,
-		IsUnixDesktopClient: isUnixRunningDesktop(),
+		IsUnixDesktopClient: util.HasGraphicalSession(),
 		Hostname:            hostName,
 		DnsLabels:           dnsLabelsReq,
 		ProfileName:         &handle,
@@ -189,7 +188,8 @@ func doExtendSession(ctx context.Context, cmd *cobra.Command) error {
 
 	client := proto.NewDaemonServiceClient(conn)
 
-	req := &proto.RequestExtendAuthSessionRequest{}
+	// the CLI runs in the user's session, the daemon does not: tell it what we can see
+	req := &proto.RequestExtendAuthSessionRequest{HasGraphicalSession: util.HasGraphicalSession()}
 	// Pre-fill the IdP login hint from the active profile so the user
 	// doesn't have to retype their email. Best-effort: we still proceed
 	// without a hint if the lookup fails.
@@ -408,9 +408,9 @@ func foregroundGetTokenInfo(ctx context.Context, cmd *cobra.Command, config *pro
 		hint = profileState.Email
 	}
 
-	oAuthFlow, err := auth.NewOAuthFlow(ctx, config, isUnixRunningDesktop(), false, hint)
+	oAuthFlow, err := auth.NewOAuthFlow(ctx, config, util.HasGraphicalSession(), false, hint)
 	if err != nil {
-		return nil, err
+		return nil, auth.WithSetupKeyAdvice(err)
 	}
 
 	flowInfo, err := oAuthFlow.RequestAuthInfo(context.TODO())
@@ -458,14 +458,6 @@ func openURL(cmd *cobra.Command, verificationURIComplete, userCode string, noBro
 	}
 }
 
-// isUnixRunningDesktop checks if a Linux OS is running desktop environment
-func isUnixRunningDesktop() bool {
-	if runtime.GOOS != "linux" && runtime.GOOS != "freebsd" {
-		return false
-	}
-	return os.Getenv("DESKTOP_SESSION") != "" || os.Getenv("XDG_CURRENT_DESKTOP") != ""
-}
-
 func setEnvAndFlags(cmd *cobra.Command) error {
 	SetFlagsFromEnvVars(rootCmd)
 

client/cmd/root.go

@@ -21,6 +21,7 @@ import (
 	"github.com/spf13/pflag"
 	"google.golang.org/grpc"
 
+	"github.com/netbirdio/netbird/client/anonymize"
 	daddr "github.com/netbirdio/netbird/client/internal/daemonaddr"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 )
@@ -69,6 +70,7 @@ var (
 	autoConnectDisabled     bool
 	extraIFaceBlackList     []string
 	anonymizeFlag           bool
+	anonymizeLevelFlag      string
 	dnsRouteInterval        time.Duration
 	// lazyConnEnabled is the parse target for the deprecated --enable-lazy-connection
 	// flag. The flag is inert; the value is no longer read (use NB_LAZY_CONN instead).
@@ -156,7 +158,8 @@ func init() {
 	rootCmd.MarkFlagsMutuallyExclusive("setup-key", "setup-key-file")
 	rootCmd.PersistentFlags().StringVar(&preSharedKey, preSharedKeyFlag, "", "Sets WireGuard PreSharedKey property. If set, then only peers that have the same key can communicate.")
 	rootCmd.PersistentFlags().StringVarP(&hostName, "hostname", "n", "", "Sets a custom hostname for the device")
-	rootCmd.PersistentFlags().BoolVarP(&anonymizeFlag, "anonymize", "A", false, "anonymize IP addresses and non-netbird.io domains in logs and status output")
+	rootCmd.PersistentFlags().BoolVarP(&anonymizeFlag, "anonymize", "A", false, "anonymize public IP addresses, MAC addresses, and non-netbird.io domains in logs and status output; private, CGNAT, and link-local IP ranges are kept (see --anonymize-level strict)")
+	rootCmd.PersistentFlags().StringVar(&anonymizeLevelFlag, "anonymize-level", "", "anonymization level: \"default\" or \"strict\"; strict also anonymizes private, CGNAT, and link-local IP ranges, peer names, and WireGuard public keys. Setting this flag implies --anonymize")
 	rootCmd.PersistentFlags().StringVarP(&configPath, "config", "c", profilemanager.DefaultConfigPath, "Overrides the default profile file location")
 
 	rootCmd.AddCommand(upCmd)
@@ -293,6 +296,19 @@ var CLIBackOffSettings = &backoff.ExponentialBackOff{
 	Clock:               backoff.SystemClock,
 }
 
+// effectiveAnonymize resolves the --anonymize and --anonymize-level flags:
+// setting a level implies anonymization, and an invalid level is rejected.
+func effectiveAnonymize() (bool, anonymize.Level, error) {
+	if anonymizeLevelFlag == "" {
+		return anonymizeFlag, anonymize.LevelDefault, nil
+	}
+	level := anonymize.ParseLevel(anonymizeLevelFlag)
+	if !strings.EqualFold(anonymizeLevelFlag, level.String()) {
+		return false, anonymize.LevelDefault, fmt.Errorf("invalid anonymize level %q: use %q or %q", anonymizeLevelFlag, anonymize.LevelDefault.String(), anonymize.LevelStrict.String())
+	}
+	return true, level, nil
+}
+
 func getSetupKey() (string, error) {
 	if setupKeyPath != "" && setupKey == "" {
 		return getSetupKeyFromFile(setupKeyPath)

client/cmd/status.go

@@ -121,8 +121,14 @@ func statusFunc(cmd *cobra.Command, args []string) error {
 		sessionExpiresAt = ts.AsTime().UTC()
 	}
 
+	anonymizeEnabled, anonymizeLevel, err := effectiveAnonymize()
+	if err != nil {
+		return err
+	}
+
 	var outputInformationHolder = nbstatus.ConvertToStatusOutputOverview(resp.GetFullStatus(), nbstatus.ConvertOptions{
-		Anonymize:            anonymizeFlag,
+		Anonymize:            anonymizeEnabled,
+		AnonymizeLevel:       anonymizeLevel,
 		DaemonVersion:        resp.GetDaemonVersion(),
 		DaemonStatus:         nbstatus.ParseDaemonStatus(status),
 		StatusFilter:         statusFilter,

client/cmd/up.go

@@ -21,8 +21,8 @@ import (
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/peer"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
-	"github.com/netbirdio/netbird/client/proto"
 	nbnet "github.com/netbirdio/netbird/client/net"
+	"github.com/netbirdio/netbird/client/proto"
 	"github.com/netbirdio/netbird/client/server"
 	"github.com/netbirdio/netbird/client/system"
 	"github.com/netbirdio/netbird/shared/management/domain"
@@ -626,7 +626,7 @@ func setupLoginRequest(providedSetupKey string, customDNSAddressConverted []byte
 		NatExternalIPs:      natExternalIPs,
 		CleanNATExternalIPs: natExternalIPs != nil && len(natExternalIPs) == 0,
 		CustomDNSAddress:    customDNSAddressConverted,
-		IsUnixDesktopClient: isUnixRunningDesktop(),
+		IsUnixDesktopClient: util.HasGraphicalSession(),
 		Hostname:            hostName,
 		ExtraIFaceBlacklist: extraIFaceBlackList,
 		DnsLabels:           dnsLabels,

client/firewall/iptables/acl_linux.go

@@ -42,6 +42,7 @@ type aclManager struct {
 	optionalEntries map[string][]entry
 	ipsetStore      *ipsetStore
 	v6              bool
+	ipsetSupported  bool
 
 	stateManager *statemanager.Manager
 }
@@ -60,6 +61,8 @@ func newAclManager(iptablesClient *iptables.IPTables, wgIface iFaceMapper) (*acl
 func (m *aclManager) init(stateManager *statemanager.Manager) error {
 	m.stateManager = stateManager
 
+	m.ipsetSupported = m.probeIPSetSupport()
+
 	m.seedInitialEntries()
 	m.seedInitialOptionalEntries()
 
@@ -91,6 +94,12 @@ func (m *aclManager) AddPeerFiltering(
 	if m.v6 && ipsetName != "" {
 		ipsetName += "-v6"
 	}
+	// When the kernel lacks the required ipset hash module, fall back to
+	// per-IP iptables rules (pre-0.68 behavior) so ACLs keep working instead
+	// of silently leaving the chain empty.
+	if ipsetName != "" && !m.ipsetSupported {
+		ipsetName = ""
+	}
 	proto := protoForFamily(protocol, m.v6)
 	specs := filterRuleSpecs(ip, proto, sPort, dPort, action, ipsetName)
 
@@ -498,6 +507,40 @@ func transformIPsetName(ipsetName string, sPort, dPort *firewall.Port, action fi
 	}
 }
 
+// probeIPSetSupport checks whether the kernel can create the ipset type used for
+// ACL rules. On kernels lacking the required ipset hash module, ipset creation
+// fails (e.g. "invalid argument"), which would otherwise leave the ACL chain
+// empty and silently drop all policy-permitted inbound traffic. When unsupported,
+// the manager falls back to per-IP iptables rules.
+func (m *aclManager) probeIPSetSupport() bool {
+	// Use a unique name so concurrent processes don't collide and we only ever
+	// destroy the set we created ourselves. ipset names are limited to 31 chars,
+	// so use a short random suffix.
+	probeName := "nb-probe-" + uuid.New().String()[:8]
+
+	opts := ipset.CreateOptions{
+		Replace: true,
+	}
+	if m.v6 {
+		opts.Family = ipset.FamilyIPV6
+	}
+
+	if err := ipset.Create(probeName, ipset.TypeHashNet, opts); err != nil {
+		log.Warnf("ipset is not available (failed to create probe set: %v); "+
+			"falling back to per-IP iptables ACL rules. Ensure the kernel provides "+
+			"the ipset hash:net module (ip_set_hash_net) for better performance with large rule sets", err)
+		return false
+	}
+
+	defer func() {
+		if err := ipset.Destroy(probeName); err != nil {
+			log.Debugf("destroy ipset probe set %q: %v", probeName, err)
+		}
+	}()
+
+	return true
+}
+
 func (m *aclManager) createIPSet(name string) error {
 	opts := ipset.CreateOptions{
 		Replace: true,

client/firewall/iptables/dnat_refcount_linux_test.go

@@ -0,0 +1,240 @@
+//go:build privileged
+
+package iptables
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	fw "github.com/netbirdio/netbird/client/firewall/manager"
+	"github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+)
+
+func iptRefcountIfaceV4() *iFaceMock {
+	return &iFaceMock{
+		NameFunc: func() string { return "wt-refcount" },
+		AddressFunc: func() wgaddr.Address {
+			return wgaddr.Address{
+				IP:      netip.MustParseAddr("10.20.0.1"),
+				Network: netip.MustParsePrefix("10.20.0.0/24"),
+			}
+		},
+	}
+}
+
+func iptRefcountIfaceDual() *iFaceMock {
+	return &iFaceMock{
+		NameFunc: func() string { return "wt-refcount" },
+		AddressFunc: func() wgaddr.Address {
+			return wgaddr.Address{
+				IP:      netip.MustParseAddr("10.20.0.1"),
+				Network: netip.MustParsePrefix("10.20.0.0/24"),
+				IPv6:    netip.MustParseAddr("fd00::1"),
+				IPv6Net: netip.MustParsePrefix("fd00::/64"),
+			}
+		},
+	}
+}
+
+func newIptRefcountManager(t *testing.T, dual bool) *Manager {
+	t.Helper()
+	var ifMock *iFaceMock
+	if dual {
+		ifMock = iptRefcountIfaceDual()
+	} else {
+		ifMock = iptRefcountIfaceV4()
+	}
+	m, err := Create(ifMock, iface.DefaultMTU)
+	require.NoError(t, err, "create manager")
+	require.NoError(t, m.Init(nil), "init manager")
+	t.Cleanup(func() {
+		require.NoError(t, m.Close(nil), "close manager")
+	})
+	return m
+}
+
+func iptDnatV4(port uint16) fw.ForwardRule {
+	return fw.ForwardRule{
+		Protocol:          fw.ProtocolTCP,
+		DestinationPort:   fw.Port{Values: []uint16{port}},
+		TranslatedAddress: netip.MustParseAddr("10.20.0.2"),
+		TranslatedPort:    fw.Port{Values: []uint16{80}},
+	}
+}
+
+func iptDnatV6(port uint16) fw.ForwardRule {
+	return fw.ForwardRule{
+		Protocol:          fw.ProtocolTCP,
+		DestinationPort:   fw.Port{Values: []uint16{port}},
+		TranslatedAddress: netip.MustParseAddr("fd00::2"),
+		TranslatedPort:    fw.Port{Values: []uint16{80}},
+	}
+}
+
+// TestIptablesRouting_RepeatedEnableSingleReference verifies that EnableRouting
+// (called on every network-map update) holds at most one reference per family
+// and a single DisableRouting drops both back to zero.
+func TestIptablesRouting_RepeatedEnableSingleReference(t *testing.T) {
+	m := newIptRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	require.NoError(t, m.EnableRouting(), "first enable")
+	require.NoError(t, m.EnableRouting(), "second enable")
+	require.NoError(t, m.EnableRouting(), "third enable")
+	v4, v6 := state.Counts()
+	assert.Equal(t, 1, v4, "repeated enable holds a single v4 reference")
+	assert.Equal(t, 1, v6, "repeated enable holds a single v6 reference")
+
+	require.NoError(t, m.DisableRouting(), "disable")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4, "single disable releases the v4 reference")
+	assert.Equal(t, 0, v6, "single disable releases the v6 reference")
+}
+
+// TestIptablesRouting_DisableKeepsDNATReference verifies that an unpaired
+// DisableRouting does not release references held by active DNAT rules.
+func TestIptablesRouting_DisableKeepsDNATReference(t *testing.T) {
+	m := newIptRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	r1, err := m.AddDNATRule(iptDnatV6(9095))
+	require.NoError(t, err, "add v6 dnat")
+
+	require.NoError(t, m.DisableRouting(), "unpaired disable")
+	_, v6 := state.Counts()
+	assert.Equal(t, 1, v6, "DNAT-held reference survives unpaired DisableRouting")
+
+	require.NoError(t, m.DeleteDNATRule(r1), "delete v6 dnat")
+	_, v6 = state.Counts()
+	assert.Equal(t, 0, v6, "delete releases the DNAT reference")
+}
+
+// TestIptablesDNAT_RefcountBalancedV4 covers a Balanced Add/Delete pair on v4.
+func TestIptablesDNAT_RefcountBalancedV4(t *testing.T) {
+	m := newIptRefcountManager(t, false)
+	state := m.router.ipFwdState
+
+	r1, err := m.AddDNATRule(iptDnatV4(7081))
+	require.NoError(t, err, "add v4 dnat 1")
+	v4, v6 := state.Counts()
+	assert.Equal(t, 1, v4, "v4 refcount after first add")
+	assert.Equal(t, 0, v6, "v6 refcount unchanged")
+
+	r2, err := m.AddDNATRule(iptDnatV4(7082))
+	require.NoError(t, err, "add v4 dnat 2")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 2, v4, "v4 refcount after second add")
+	assert.Equal(t, 0, v6, "v6 refcount unchanged")
+
+	require.NoError(t, m.DeleteDNATRule(r1))
+	v4, v6 = state.Counts()
+	assert.Equal(t, 1, v4, "v4 refcount after first delete")
+	assert.Equal(t, 0, v6, "v6 refcount unchanged")
+
+	require.NoError(t, m.DeleteDNATRule(r2))
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4, "v4 refcount after second delete")
+	assert.Equal(t, 0, v6, "v6 refcount unchanged")
+}
+
+// TestIptablesDNAT_RefcountBalancedV6 checks the v6 path increments v6 only and
+// decrements back to zero.
+func TestIptablesDNAT_RefcountBalancedV6(t *testing.T) {
+	m := newIptRefcountManager(t, true)
+	require.NotNil(t, m.router6, "v6 router")
+	require.Same(t, m.router.ipFwdState, m.router6.ipFwdState, "shared state")
+	state := m.router.ipFwdState
+
+	r1, err := m.AddDNATRule(iptDnatV6(9081))
+	require.NoError(t, err, "add v6 dnat 1")
+	v4, v6 := state.Counts()
+	assert.Equal(t, 0, v4)
+	assert.Equal(t, 1, v6, "v6 refcount after first add")
+
+	r2, err := m.AddDNATRule(iptDnatV6(9082))
+	require.NoError(t, err, "add v6 dnat 2")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4, "v4 refcount unchanged")
+	assert.Equal(t, 2, v6, "v6 refcount after second add")
+
+	require.NoError(t, m.DeleteDNATRule(r1))
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4, "v4 refcount unchanged")
+	assert.Equal(t, 1, v6, "v6 refcount after first delete")
+
+	require.NoError(t, m.DeleteDNATRule(r2))
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4)
+	assert.Equal(t, 0, v6, "v6 refcount after second delete")
+}
+
+// TestIptablesDNAT_DuplicateAddNoLeak verifies the duplicate-rule path returns
+// without bumping the refcount.
+func TestIptablesDNAT_DuplicateAddNoLeak(t *testing.T) {
+	m := newIptRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	rule := iptDnatV4(7083)
+	r1, err := m.AddDNATRule(rule)
+	require.NoError(t, err)
+	v4, _ := state.Counts()
+	assert.Equal(t, 1, v4)
+
+	_, err = m.AddDNATRule(rule)
+	require.NoError(t, err, "duplicate add")
+	v4, _ = state.Counts()
+	assert.Equal(t, 1, v4, "duplicate add must not increment")
+
+	require.NoError(t, m.DeleteDNATRule(r1))
+	v4, _ = state.Counts()
+	assert.Equal(t, 0, v4, "single delete must drop to zero")
+}
+
+// TestIptablesDNAT_DeleteMissingNoUnderflow verifies Delete on an unknown rule
+// neither errors nor releases the refcount.
+func TestIptablesDNAT_DeleteMissingNoUnderflow(t *testing.T) {
+	m := newIptRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	phantom := iptDnatV4(7099)
+	require.NoError(t, m.DeleteDNATRule(&phantom), "delete missing v4")
+	v4, v6 := state.Counts()
+	assert.Equal(t, 0, v4)
+	assert.Equal(t, 0, v6)
+
+	phantom6 := iptDnatV6(9099)
+	require.NoError(t, m.DeleteDNATRule(&phantom6), "delete missing v6")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4)
+	assert.Equal(t, 0, v6)
+
+	r1, err := m.AddDNATRule(iptDnatV4(7100))
+	require.NoError(t, err)
+	v4, _ = state.Counts()
+	assert.Equal(t, 1, v4, "real add still increments after phantom delete")
+	require.NoError(t, m.DeleteDNATRule(r1))
+}
+
+// TestIptablesDNAT_DoubleDeleteNoUnderflow verifies a second Delete on the same
+// rule is a no-op.
+func TestIptablesDNAT_DoubleDeleteNoUnderflow(t *testing.T) {
+	m := newIptRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	r1, err := m.AddDNATRule(iptDnatV6(9083))
+	require.NoError(t, err)
+	_, v6 := state.Counts()
+	assert.Equal(t, 1, v6)
+
+	require.NoError(t, m.DeleteDNATRule(r1), "first delete")
+	_, v6 = state.Counts()
+	assert.Equal(t, 0, v6)
+
+	require.NoError(t, m.DeleteDNATRule(r1), "second delete must be no-op")
+	_, v6 = state.Counts()
+	assert.Equal(t, 0, v6, "double delete must not underflow")
+}

client/firewall/iptables/manager_linux.go

@@ -89,7 +89,7 @@ func (m *Manager) createIPv6Components(wgIface iFaceMapper, mtu uint16) error {
 	}
 
 	// Share the same IP forwarding state with the v4 router, since
-	// EnableIPForwarding controls both v4 and v6 sysctls.
+	// Forwarding refcounter is per-family but shared between v4 and v6 routers.
 	m.router6.ipFwdState = m.router.ipFwdState
 
 	m.aclMgr6, err = newAclManager(ip6Client, wgIface)
@@ -402,17 +402,12 @@ func (m *Manager) SetLogLevel(log.Level) {
 }
 
 func (m *Manager) EnableRouting() error {
-	if err := m.router.ipFwdState.RequestForwarding(); err != nil {
-		return fmt.Errorf("enable IP forwarding: %w", err)
-	}
-	return nil
+	// v6 only when the overlay actually has v6.
+	return m.router.ipFwdState.RequestRouting(m.router6 != nil)
 }
 
 func (m *Manager) DisableRouting() error {
-	if err := m.router.ipFwdState.ReleaseForwarding(); err != nil {
-		return fmt.Errorf("disable IP forwarding: %w", err)
-	}
-	return nil
+	return m.router.ipFwdState.ReleaseRouting()
 }
 
 // AddDNATRule adds a DNAT rule

client/firewall/iptables/manager_linux_test.go

@@ -291,3 +291,40 @@ func TestIptablesCreatePerformance(t *testing.T) {
 		})
 	}
 }
+
+// TestIptablesACLIPSetFallback verifies that when the kernel lacks ipset support,
+// the ACL manager falls back to per-IP iptables rules (-s <ip>) instead of
+// silently leaving the chain empty. See discussion #6125.
+func TestIptablesACLIPSetFallback(t *testing.T) {
+	ipv4Client, err := iptables.NewWithProtocol(iptables.ProtocolIPv4)
+	require.NoError(t, err)
+
+	// Use Create()/Init() so the router-owned chains (chainRTFWDIN/OUT) are
+	// created before the ACL manager's createDefaultChains() references them.
+	manager, err := Create(ifaceMock, iface.DefaultMTU)
+	require.NoError(t, err)
+	require.NoError(t, manager.Init(nil))
+
+	aclMgr := manager.aclMgr
+	// Simulate a kernel without the ipset hash module.
+	aclMgr.ipsetSupported = false
+
+	defer func() {
+		require.NoError(t, manager.Close(nil))
+	}()
+
+	ip := netip.MustParseAddr("10.20.0.42")
+	port := &fw.Port{Values: []uint16{22}}
+
+	rules, err := aclMgr.AddPeerFiltering(nil, ip.AsSlice(), "tcp", nil, port, fw.ActionAccept, "nb0000001")
+	require.NoError(t, err, "AddPeerFiltering should succeed via fallback")
+	require.NotEmpty(t, rules)
+
+	rule := rules[0].(*Rule)
+	require.Empty(t, rule.ipsetName, "fallback rule must not reference an ipset")
+	require.Contains(t, strings.Join(rule.specs, " "), "-s 10.20.0.42", "fallback rule must match by source IP")
+	require.NotContains(t, strings.Join(rule.specs, " "), "--match-set", "fallback rule must not use ipset matching")
+
+	// The rule must actually be present in the ACL chain (not silently dropped).
+	checkRuleSpecs(t, ipv4Client, rule.chain, true, rule.specs...)
+}

client/firewall/iptables/router_linux.go

@@ -102,7 +102,7 @@ func newRouter(iptablesClient *iptables.IPTables, wgIface iFaceMapper, mtu uint1
 		wgIface:        wgIface,
 		mtu:            mtu,
 		v6:             iptablesClient.Proto() == iptables.ProtocolIPv6,
-		ipFwdState:     ipfwdstate.NewIPForwardingState(),
+		ipFwdState:     ipfwdstate.NewIPForwardingState(wgIface.Name()),
 	}
 
 	r.ipsetCounter = refcounter.New(
@@ -770,10 +770,6 @@ func (r *router) updateState() {
 }
 
 func (r *router) AddDNATRule(rule firewall.ForwardRule) (firewall.Rule, error) {
-	if err := r.ipFwdState.RequestForwarding(); err != nil {
-		return nil, err
-	}
-
 	ruleKey := rule.ID()
 	if _, exists := r.rules[ruleKey+dnatSuffix]; exists {
 		return rule, nil
@@ -840,18 +836,34 @@ func (r *router) AddDNATRule(rule firewall.ForwardRule) (firewall.Rule, error) {
 
 	for key, ruleInfo := range rules {
 		if err := r.iptablesClient.Append(ruleInfo.table, ruleInfo.chain, ruleInfo.rule...); err != nil {
-			if rollbackErr := r.rollbackRules(rules); rollbackErr != nil {
-				log.Errorf("rollback failed: %v", rollbackErr)
-			}
+			r.cleanupFailedDNATAdd(rules)
 			return nil, fmt.Errorf("add rule %s: %w", key, err)
 		}
 		r.rules[key] = ruleInfo.rule
 	}
 
+	if err := r.ipFwdState.RequestForwarding(r.v6); err != nil {
+		r.cleanupFailedDNATAdd(rules)
+		return nil, fmt.Errorf("enable forwarding: %w", err)
+	}
+
 	r.updateState()
 	return rule, nil
 }
 
+// cleanupFailedDNATAdd removes the bookkeeping written by a partially applied
+// AddDNATRule before rolling back the kernel rules, so no entries remain that
+// never got a forwarding refcount. rollbackRules re-adds entries it failed to
+// remove from the kernel.
+func (r *router) cleanupFailedDNATAdd(rules map[string]ruleInfo) {
+	for key := range rules {
+		delete(r.rules, key)
+	}
+	if err := r.rollbackRules(rules); err != nil {
+		log.Errorf("rollback failed: %v", err)
+	}
+}
+
 func (r *router) rollbackRules(rules map[string]ruleInfo) error {
 	var merr *multierror.Error
 	for key, ruleInfo := range rules {
@@ -868,32 +880,47 @@ func (r *router) rollbackRules(rules map[string]ruleInfo) error {
 }
 
 func (r *router) DeleteDNATRule(rule firewall.Rule) error {
-	if err := r.ipFwdState.ReleaseForwarding(); err != nil {
-		log.Errorf("%v", err)
-	}
-
 	ruleKey := rule.ID()
 
+	_, hadDNAT := r.rules[ruleKey+dnatSuffix]
+	_, hadSNAT := r.rules[ruleKey+snatSuffix]
+	_, hadFWD := r.rules[ruleKey+fwdSuffix]
+	if !hadDNAT && !hadSNAT && !hadFWD {
+		return nil
+	}
+
 	var merr *multierror.Error
 	if dnatRule, exists := r.rules[ruleKey+dnatSuffix]; exists {
 		if err := r.iptablesClient.Delete(tableNat, chainRTRDR, dnatRule...); err != nil {
 			merr = multierror.Append(merr, fmt.Errorf("delete DNAT rule: %w", err))
+		} else {
+			delete(r.rules, ruleKey+dnatSuffix)
 		}
-		delete(r.rules, ruleKey+dnatSuffix)
 	}
 
 	if snatRule, exists := r.rules[ruleKey+snatSuffix]; exists {
 		if err := r.iptablesClient.Delete(tableNat, chainRTNAT, snatRule...); err != nil {
 			merr = multierror.Append(merr, fmt.Errorf("delete SNAT rule: %w", err))
+		} else {
+			delete(r.rules, ruleKey+snatSuffix)
 		}
-		delete(r.rules, ruleKey+snatSuffix)
 	}
 
 	if fwdRule, exists := r.rules[ruleKey+fwdSuffix]; exists {
 		if err := r.iptablesClient.Delete(tableFilter, chainRTFWDOUT, fwdRule...); err != nil {
 			merr = multierror.Append(merr, fmt.Errorf("delete forward rule: %w", err))
+		} else {
+			delete(r.rules, ruleKey+fwdSuffix)
+		}
+	}
+
+	// Release the refcount only once all rules are gone from the kernel. On
+	// partial failure the failed entries stay in r.rules so a retry can remove
+	// them and release then.
+	if merr == nil {
+		if err := r.ipFwdState.ReleaseForwarding(r.v6); err != nil {
+			log.Errorf("%v", err)
 		}
-		delete(r.rules, ruleKey+fwdSuffix)
 	}
 
 	r.updateState()

client/firewall/nftables/dnat_refcount_linux_test.go

@@ -0,0 +1,249 @@
+//go:build privileged
+
+package nftables
+
+import (
+	"net/netip"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+
+	fw "github.com/netbirdio/netbird/client/firewall/manager"
+	"github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/iface/wgaddr"
+)
+
+func nftRefcountIfaceV4() *iFaceMock {
+	return &iFaceMock{
+		NameFunc: func() string { return "wt-refcount" },
+		AddressFunc: func() wgaddr.Address {
+			return wgaddr.Address{
+				IP:      netip.MustParseAddr("100.96.0.1"),
+				Network: netip.MustParsePrefix("100.96.0.0/16"),
+			}
+		},
+	}
+}
+
+func nftRefcountIfaceDual() *iFaceMock {
+	return &iFaceMock{
+		NameFunc: func() string { return "wt-refcount" },
+		AddressFunc: func() wgaddr.Address {
+			return wgaddr.Address{
+				IP:      netip.MustParseAddr("100.96.0.1"),
+				Network: netip.MustParsePrefix("100.96.0.0/16"),
+				IPv6:    netip.MustParseAddr("fd00::1"),
+				IPv6Net: netip.MustParsePrefix("fd00::/64"),
+			}
+		},
+	}
+}
+
+func newNftRefcountManager(t *testing.T, dual bool) *Manager {
+	t.Helper()
+	if check() != NFTABLES {
+		t.Skip("nftables not supported on this system")
+	}
+	var ifMock *iFaceMock
+	if dual {
+		ifMock = nftRefcountIfaceDual()
+	} else {
+		ifMock = nftRefcountIfaceV4()
+	}
+	m, err := Create(ifMock, iface.DefaultMTU)
+	require.NoError(t, err, "create manager")
+	require.NoError(t, m.Init(nil), "init manager")
+	t.Cleanup(func() {
+		require.NoError(t, m.Close(nil), "close manager")
+	})
+	return m
+}
+
+func dnatV4(port uint16) fw.ForwardRule {
+	return fw.ForwardRule{
+		Protocol:          fw.ProtocolTCP,
+		DestinationPort:   fw.Port{Values: []uint16{port}},
+		TranslatedAddress: netip.MustParseAddr("100.96.0.2"),
+		TranslatedPort:    fw.Port{Values: []uint16{80}},
+	}
+}
+
+func dnatV6(port uint16) fw.ForwardRule {
+	return fw.ForwardRule{
+		Protocol:          fw.ProtocolTCP,
+		DestinationPort:   fw.Port{Values: []uint16{port}},
+		TranslatedAddress: netip.MustParseAddr("fd00::2"),
+		TranslatedPort:    fw.Port{Values: []uint16{80}},
+	}
+}
+
+// TestNftablesDNAT_RefcountBalancedV4 verifies that Add/Delete pairs leave the
+// v4 refcount at zero.
+func TestNftablesDNAT_RefcountBalancedV4(t *testing.T) {
+	m := newNftRefcountManager(t, false)
+	state := m.router.ipFwdState
+
+	r1, err := m.AddDNATRule(dnatV4(8081))
+	require.NoError(t, err, "add v4 dnat 1")
+	v4, v6 := state.Counts()
+	assert.Equal(t, 1, v4, "v4 refcount after first add")
+	assert.Equal(t, 0, v6, "v6 refcount unchanged")
+
+	r2, err := m.AddDNATRule(dnatV4(8082))
+	require.NoError(t, err, "add v4 dnat 2")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 2, v4, "v4 refcount after second add")
+	assert.Equal(t, 0, v6, "v6 refcount unchanged")
+
+	require.NoError(t, m.DeleteDNATRule(r1), "delete v4 dnat 1")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 1, v4, "v4 refcount after first delete")
+	assert.Equal(t, 0, v6, "v6 refcount unchanged")
+
+	require.NoError(t, m.DeleteDNATRule(r2), "delete v4 dnat 2")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4, "v4 refcount after second delete")
+	assert.Equal(t, 0, v6, "v6 refcount unchanged")
+}
+
+// TestNftablesDNAT_RefcountBalancedV6 verifies the v6 path increments v6 only
+// and decrements back to zero on Delete.
+func TestNftablesDNAT_RefcountBalancedV6(t *testing.T) {
+	m := newNftRefcountManager(t, true)
+	require.NotNil(t, m.router6, "v6 router")
+	require.Same(t, m.router.ipFwdState, m.router6.ipFwdState, "shared state")
+	state := m.router.ipFwdState
+
+	r1, err := m.AddDNATRule(dnatV6(9091))
+	require.NoError(t, err, "add v6 dnat 1")
+	v4, v6 := state.Counts()
+	assert.Equal(t, 0, v4, "v4 refcount unchanged")
+	assert.Equal(t, 1, v6, "v6 refcount after first add")
+
+	r2, err := m.AddDNATRule(dnatV6(9092))
+	require.NoError(t, err, "add v6 dnat 2")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4)
+	assert.Equal(t, 2, v6, "v6 refcount after second add")
+
+	require.NoError(t, m.DeleteDNATRule(r1), "delete v6 dnat 1")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4, "v4 refcount unchanged")
+	assert.Equal(t, 1, v6, "v6 refcount after first delete")
+
+	require.NoError(t, m.DeleteDNATRule(r2), "delete v6 dnat 2")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4)
+	assert.Equal(t, 0, v6, "v6 refcount after second delete")
+}
+
+// TestNftablesDNAT_DuplicateAddNoLeak verifies that a duplicate Add (same
+// ForwardRule) does not double-increment the refcount.
+func TestNftablesDNAT_DuplicateAddNoLeak(t *testing.T) {
+	m := newNftRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	rule := dnatV4(8083)
+	r1, err := m.AddDNATRule(rule)
+	require.NoError(t, err, "add v4 dnat")
+	v4, _ := state.Counts()
+	assert.Equal(t, 1, v4)
+
+	// duplicate add: same rule ID, must be a no-op for the refcount.
+	_, err = m.AddDNATRule(rule)
+	require.NoError(t, err, "duplicate add")
+	v4, _ = state.Counts()
+	assert.Equal(t, 1, v4, "duplicate add must not increment")
+
+	require.NoError(t, m.DeleteDNATRule(r1), "delete v4 dnat")
+	v4, _ = state.Counts()
+	assert.Equal(t, 0, v4, "single delete must drop to zero")
+}
+
+// TestNftablesDNAT_DeleteMissingNoUnderflow verifies deleting a rule that was
+// never added does not underflow the refcount.
+func TestNftablesDNAT_DeleteMissingNoUnderflow(t *testing.T) {
+	m := newNftRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	// Construct a Rule reference for something never added. The router stores
+	// rules by ID(), and DeleteDNATRule looks them up in r.rules; a missing
+	// entry must be a no-op rather than calling Release.
+	phantom := dnatV4(8099)
+	require.NoError(t, m.DeleteDNATRule(&phantom), "delete missing v4 dnat")
+	v4, v6 := state.Counts()
+	assert.Equal(t, 0, v4, "v4 refcount unaffected by missing delete")
+	assert.Equal(t, 0, v6, "v6 refcount unaffected")
+
+	phantom6 := dnatV6(9099)
+	require.NoError(t, m.DeleteDNATRule(&phantom6), "delete missing v6 dnat")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4)
+	assert.Equal(t, 0, v6, "v6 refcount unaffected by missing delete")
+
+	// And after a phantom delete, a real add still results in count=1.
+	r1, err := m.AddDNATRule(dnatV4(8100))
+	require.NoError(t, err, "add v4 dnat after phantom delete")
+	v4, _ = state.Counts()
+	assert.Equal(t, 1, v4, "real add still increments after phantom delete")
+	require.NoError(t, m.DeleteDNATRule(r1))
+}
+
+// TestNftablesRouting_RepeatedEnableSingleReference verifies that EnableRouting
+// (called on every network-map update) holds at most one reference per family
+// and a single DisableRouting drops both back to zero.
+func TestNftablesRouting_RepeatedEnableSingleReference(t *testing.T) {
+	m := newNftRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	require.NoError(t, m.EnableRouting(), "first enable")
+	require.NoError(t, m.EnableRouting(), "second enable")
+	require.NoError(t, m.EnableRouting(), "third enable")
+	v4, v6 := state.Counts()
+	assert.Equal(t, 1, v4, "repeated enable holds a single v4 reference")
+	assert.Equal(t, 1, v6, "repeated enable holds a single v6 reference")
+
+	require.NoError(t, m.DisableRouting(), "disable")
+	v4, v6 = state.Counts()
+	assert.Equal(t, 0, v4, "single disable releases the v4 reference")
+	assert.Equal(t, 0, v6, "single disable releases the v6 reference")
+}
+
+// TestNftablesRouting_DisableKeepsDNATReference verifies that an unpaired
+// DisableRouting does not release references held by active DNAT rules.
+func TestNftablesRouting_DisableKeepsDNATReference(t *testing.T) {
+	m := newNftRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	r1, err := m.AddDNATRule(dnatV6(9095))
+	require.NoError(t, err, "add v6 dnat")
+
+	require.NoError(t, m.DisableRouting(), "unpaired disable")
+	_, v6 := state.Counts()
+	assert.Equal(t, 1, v6, "DNAT-held reference survives unpaired DisableRouting")
+
+	require.NoError(t, m.DeleteDNATRule(r1), "delete v6 dnat")
+	_, v6 = state.Counts()
+	assert.Equal(t, 0, v6, "delete releases the DNAT reference")
+}
+
+// TestNftablesDNAT_DoubleDeleteNoUnderflow verifies that deleting the same rule
+// twice does not underflow the refcount (the second delete is a no-op).
+func TestNftablesDNAT_DoubleDeleteNoUnderflow(t *testing.T) {
+	m := newNftRefcountManager(t, true)
+	state := m.router.ipFwdState
+
+	r1, err := m.AddDNATRule(dnatV6(9093))
+	require.NoError(t, err)
+	_, v6 := state.Counts()
+	assert.Equal(t, 1, v6)
+
+	require.NoError(t, m.DeleteDNATRule(r1), "first delete")
+	_, v6 = state.Counts()
+	assert.Equal(t, 0, v6)
+
+	require.NoError(t, m.DeleteDNATRule(r1), "second delete must be no-op")
+	_, v6 = state.Counts()
+	assert.Equal(t, 0, v6, "double delete must not underflow")
+}

client/firewall/nftables/manager_linux.go

@@ -105,8 +105,8 @@ func (m *Manager) createIPv6Components(tableName string, wgIface iFaceMapper, mt
 		return fmt.Errorf("create v6 router: %w", err)
 	}
 
-	// Share the same IP forwarding state with the v4 router, since
-	// EnableIPForwarding controls both v4 and v6 sysctls.
+	// Share the per-family forwarding refcounter with the v4 router so a v4
+	// rule and a v6 rule against the same state machine cooperate cleanly.
 	m.router6.ipFwdState = m.router.ipFwdState
 
 	m.aclManager6, err = newAclManager(workTable6, wgIface, chainNameRoutingFw)
@@ -530,17 +530,12 @@ func (m *Manager) SetLogLevel(log.Level) {
 }
 
 func (m *Manager) EnableRouting() error {
-	if err := m.router.ipFwdState.RequestForwarding(); err != nil {
-		return fmt.Errorf("enable IP forwarding: %w", err)
-	}
-	return nil
+	// v6 only when the overlay actually has v6.
+	return m.router.ipFwdState.RequestRouting(m.router6 != nil)
 }
 
 func (m *Manager) DisableRouting() error {
-	if err := m.router.ipFwdState.ReleaseForwarding(); err != nil {
-		return fmt.Errorf("disable IP forwarding: %w", err)
-	}
-	return nil
+	return m.router.ipFwdState.ReleaseRouting()
 }
 
 // Flush rule/chain/set operations from the buffer

client/firewall/nftables/router_linux.go

@@ -93,7 +93,7 @@ func newRouter(workTable *nftables.Table, wgIface iFaceMapper, mtu uint16) (*rou
 		rules:      make(map[string]*nftables.Rule),
 		af:         familyForAddr(workTable.Family == nftables.TableFamilyIPv4),
 		wgIface:    wgIface,
-		ipFwdState: ipfwdstate.NewIPForwardingState(),
+		ipFwdState: ipfwdstate.NewIPForwardingState(wgIface.Name()),
 		mtu:        mtu,
 	}
 
@@ -1553,10 +1553,6 @@ func (r *router) refreshRulesMap() error {
 }
 
 func (r *router) AddDNATRule(rule firewall.ForwardRule) (firewall.Rule, error) {
-	if err := r.ipFwdState.RequestForwarding(); err != nil {
-		return nil, err
-	}
-
 	ruleKey := rule.ID()
 	if _, exists := r.rules[ruleKey+dnatSuffix]; exists {
 		return rule, nil
@@ -1567,7 +1563,18 @@ func (r *router) AddDNATRule(rule firewall.ForwardRule) (firewall.Rule, error) {
 		return nil, fmt.Errorf("convert protocol to number: %w", err)
 	}
 
+	// Request forwarding before queueing rules: addDnatRedirect/addDnatMasq
+	// buffer netlink messages on r.conn that the next caller's Flush would
+	// commit if we returned without flushing them ourselves.
+	v6 := r.af.tableFamily == nftables.TableFamilyIPv6
+	if err := r.ipFwdState.RequestForwarding(v6); err != nil {
+		return nil, fmt.Errorf("enable forwarding: %w", err)
+	}
+
 	if err := r.addDnatRedirect(rule, protoNum, ruleKey); err != nil {
+		if rerr := r.ipFwdState.ReleaseForwarding(v6); rerr != nil {
+			log.Warnf("rollback forwarding refcount: %v", rerr)
+		}
 		return nil, err
 	}
 
@@ -1579,6 +1586,11 @@ func (r *router) AddDNATRule(rule firewall.ForwardRule) (firewall.Rule, error) {
 	// TODO: find chains with drop policies and add rules there
 
 	if err := r.conn.Flush(); err != nil {
+		if rerr := r.ipFwdState.ReleaseForwarding(v6); rerr != nil {
+			log.Warnf("rollback forwarding refcount: %v", rerr)
+		}
+		delete(r.rules, ruleKey+dnatSuffix)
+		delete(r.rules, ruleKey+snatSuffix)
 		return nil, fmt.Errorf("flush rules: %w", err)
 	}
 
@@ -1781,16 +1793,18 @@ func (r *router) addDnatMasq(rule firewall.ForwardRule, protoNum uint8, ruleKey
 }
 
 func (r *router) DeleteDNATRule(rule firewall.Rule) error {
-	if err := r.ipFwdState.ReleaseForwarding(); err != nil {
-		log.Errorf("%v", err)
-	}
-
 	ruleKey := rule.ID()
 
 	if err := r.refreshRulesMap(); err != nil {
 		return fmt.Errorf(refreshRulesMapError, err)
 	}
 
+	_, hadDNAT := r.rules[ruleKey+dnatSuffix]
+	_, hadSNAT := r.rules[ruleKey+snatSuffix]
+	if !hadDNAT && !hadSNAT {
+		return nil
+	}
+
 	var merr *multierror.Error
 	var needsFlush bool
 
@@ -1822,9 +1836,16 @@ func (r *router) DeleteDNATRule(rule firewall.Rule) error {
 		}
 	}
 
+	// Release the refcount only once the rules are gone from the kernel. On
+	// failure (including the refreshRulesMap error above) the rules and their
+	// map entries remain, keeping forwarding on until a retry removes them.
 	if merr == nil {
 		delete(r.rules, ruleKey+dnatSuffix)
 		delete(r.rules, ruleKey+snatSuffix)
+
+		if err := r.ipFwdState.ReleaseForwarding(r.af.tableFamily == nftables.TableFamilyIPv6); err != nil {
+			log.Errorf("%v", err)
+		}
 	}
 
 	return nberrors.FormatErrorOrNil(merr)

client/iface/bind/ice_bind.go

@@ -22,6 +22,16 @@ import (
 	nbnet "github.com/netbirdio/netbird/client/net"
 )
 
+const (
+	// wgMsgTypeHandshakeInitiation is the lowest WireGuard message type.
+	wgMsgTypeHandshakeInitiation uint32 = 1
+	// wgMsgTypeTransport is the highest WireGuard message type.
+	wgMsgTypeTransport uint32 = 4
+	// wgMinMsgSize is the smallest WireGuard message: transport data with an empty
+	// payload, which is what a keepalive is.
+	wgMinMsgSize = 32
+)
+
 type receiverCreator struct {
 	iceBind *ICEBind
 }
@@ -216,8 +226,15 @@ func (s *ICEBind) createReceiverFn(pc wgConn.BatchReader, conn *net.UDPConn, rxO
 		for i := 0; i < numMsgs; i++ {
 			msg := &(*msgs)[i]
 
-			// todo: handle err
-			if ok, _ := s.filterOutStunMessages(msg.Buffers, msg.N, msg.Addr); ok {
+			if ok, err := s.filterOutStunMessages(msg.Buffers, msg.N, msg.Addr); ok {
+				if err != nil {
+					log.Debugf("failed to handle STUN packet from %s: %v", msg.Addr, err)
+				}
+				// WireGuard reuses sizes and eps across reads and only skips a slot
+				// whose size is below the minimum message size. Leaving a consumed
+				// slot untouched makes it process this buffer again under the
+				// previous packet's length and endpoint.
+				sizes[i] = 0
 				continue
 			}
 			sizes[i] = msg.N
@@ -271,11 +288,16 @@ func (s *ICEBind) createOrUpdateMux() {
 
 func (s *ICEBind) filterOutStunMessages(buffers [][]byte, n int, addr net.Addr) (bool, error) {
 	for i := range buffers {
-		if !stun.IsMessage(buffers[i]) {
+		if n > len(buffers[i]) {
+			continue
+		}
+		pkt := buffers[i][:n]
+
+		if isWireGuardMsg(pkt) || !stun.IsMessage(pkt) {
 			continue
 		}
 
-		msg, err := s.parseSTUNMessage(buffers[i][:n])
+		msg, err := s.parseSTUNMessage(pkt)
 		if err != nil {
 			buffers[i] = []byte{}
 			return true, err
@@ -347,18 +369,34 @@ func putMessages(msgs *[]ipv6.Message, msgsPool *sync.Pool) {
 	msgsPool.Put(msgs)
 }
 
-func isTransportPkg(buffers [][]byte, n int) bool {
-	// The first buffer should contain at least 4 bytes for type
-	if len(buffers[0]) < 4 {
-		return true
+// isWireGuardMsg reports whether the packet carries a WireGuard message header: a
+// little-endian uint32 message type in the range 1..4, which leaves the three bytes
+// after the type byte zero, in a packet long enough to hold any WireGuard message.
+//
+// A well formed STUN message cannot take that shape. Its length field sits in the two
+// bytes the type must leave zero, and for a message of at least wgMinMsgSize bytes that
+// field holds at least 12, so the two framings do not overlap. The test has to be this
+// tight because stun.IsMessage only looks at the magic cookie, which in a WireGuard
+// message overlaps the receiver index: a session whose index happens to equal the cookie
+// would otherwise have all of its inbound data misrouted to the STUN handler until the
+// next rekey.
+func isWireGuardMsg(pkt []byte) bool {
+	if len(pkt) < wgMinMsgSize {
+		return false
 	}
 
-	// WireGuard packet type is a little-endian uint32 at start
-	packetType := binary.LittleEndian.Uint32(buffers[0][:4])
-
-	// Check if packetType matches known WireGuard message types
-	if packetType == 4 && n > 32 {
-		return true
-	}
-	return false
+	msgType := binary.LittleEndian.Uint32(pkt[:4])
+	return msgType >= wgMsgTypeHandshakeInitiation && msgType <= wgMsgTypeTransport
+}
+
+// isTransportPkg reports whether the packet is WireGuard transport data carrying a
+// payload, which is what counts as peer activity. A keepalive holds no payload and is
+// exactly wgMinMsgSize bytes.
+func isTransportPkg(buffers [][]byte, n int) bool {
+	if n < 4 || n > len(buffers[0]) {
+		return false
+	}
+
+	msgType := binary.LittleEndian.Uint32(buffers[0][:4])
+	return msgType == wgMsgTypeTransport && n > wgMinMsgSize
 }

client/iface/bind/stun_filter_test.go

@@ -0,0 +1,215 @@
+//go:build !js
+
+package bind
+
+import (
+	"encoding/binary"
+	"net"
+	"testing"
+	"time"
+
+	"github.com/pion/stun/v3"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/net/ipv4"
+	wgConn "golang.zx2c4.com/wireguard/conn"
+)
+
+// magicCookieBytes is the STUN magic cookie as it appears on the wire. In a
+// WireGuard message the same offset holds the receiver (or sender) index, which is
+// a random uint32, so a session can draw exactly this value.
+var magicCookieBytes = []byte{0x21, 0x12, 0xA4, 0x42}
+
+const testBufSize = 1500
+
+// wgMsg builds a WireGuard message of the given type and size, with the index field
+// at bytes 4:8 set to index.
+func wgMsg(msgType uint32, size int, index []byte) []byte {
+	pkt := make([]byte, size)
+	binary.LittleEndian.PutUint32(pkt[:4], msgType)
+	copy(pkt[4:8], index)
+	return pkt
+}
+
+// intoBuffer copies pkt into a full-size receive buffer, the way the kernel read
+// does, so tests see the same buffer/length split as the hot path.
+func intoBuffer(pkt []byte) [][]byte {
+	buf := make([]byte, testBufSize)
+	copy(buf, pkt)
+	return [][]byte{buf}
+}
+
+func TestFilterOutStunMessages_PassesWireGuardWithCookieShapedIndex(t *testing.T) {
+	tests := []struct {
+		name    string
+		msgType uint32
+		size    int
+	}{
+		{"transport data", wgMsgTypeTransport, 128},
+		{"keepalive", wgMsgTypeTransport, wgMinMsgSize},
+		{"handshake initiation", wgMsgTypeHandshakeInitiation, 148},
+		{"handshake response", 2, 92},
+		{"cookie reply", 3, 64},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			pkt := wgMsg(tc.msgType, tc.size, magicCookieBytes)
+			require.True(t, stun.IsMessage(pkt), "precondition: pion sees this as STUN")
+
+			buffers := intoBuffer(pkt)
+			bind := &ICEBind{}
+
+			filtered, err := bind.filterOutStunMessages(buffers, tc.size, &net.UDPAddr{})
+			assert.NoError(t, err)
+			assert.False(t, filtered, "WireGuard message must be handed to WireGuard, not the STUN handler")
+			assert.Len(t, buffers[0], testBufSize, "buffer must be left intact for WireGuard")
+		})
+	}
+}
+
+func TestFilterOutStunMessages_FiltersRealSTUNMessage(t *testing.T) {
+	msg, err := stun.Build(stun.BindingRequest, stun.TransactionID, stun.Fingerprint)
+	require.NoError(t, err)
+
+	buffers := intoBuffer(msg.Raw)
+	bind := &ICEBind{}
+
+	filtered, err := bind.filterOutStunMessages(buffers, len(msg.Raw), &net.UDPAddr{})
+	assert.NoError(t, err)
+	assert.True(t, filtered, "STUN message must be consumed by the STUN handler")
+	assert.Empty(t, buffers[0], "consumed buffer must be emptied so WireGuard does not see it")
+}
+
+// TestIsWireGuardMsg_DisjointFromSTUN locks the invariant the filter relies on: a
+// well formed STUN message long enough to be a WireGuard message always has a
+// non-zero length field, so it cannot be mistaken for a WireGuard header.
+func TestIsWireGuardMsg_DisjointFromSTUN(t *testing.T) {
+	types := []stun.MessageType{
+		stun.BindingRequest,
+		stun.BindingSuccess,
+		stun.BindingError,
+		{Method: stun.MethodBinding, Class: stun.ClassIndication},
+	}
+
+	for _, msgType := range types {
+		// Long enough that the length guard is not what makes this pass.
+		msg, err := stun.Build(msgType, stun.TransactionID,
+			stun.NewUsername("remoteUfrag:localUfrag"), stun.Fingerprint)
+		require.NoError(t, err)
+		require.GreaterOrEqual(t, len(msg.Raw), wgMinMsgSize, "precondition: %s", msgType)
+		assert.False(t, isWireGuardMsg(msg.Raw),
+			"%s must not look like a WireGuard message", msgType)
+	}
+}
+
+func TestIsWireGuardMsg(t *testing.T) {
+	tests := []struct {
+		name string
+		pkt  []byte
+		want bool
+	}{
+		{"transport data", wgMsg(wgMsgTypeTransport, 128, nil), true},
+		{"handshake initiation", wgMsg(wgMsgTypeHandshakeInitiation, 148, nil), true},
+		{"unknown type 5", wgMsg(5, 128, nil), false},
+		{"type 0", wgMsg(0, 128, nil), false},
+		{"non-zero reserved byte", []byte{0x04, 0x00, 0x01, 0x00}, false},
+		{"too short", []byte{0x04, 0x00, 0x00}, false},
+		{"empty", nil, false},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			assert.Equal(t, tc.want, isWireGuardMsg(tc.pkt), "wrong classification for %s", tc.name)
+		})
+	}
+}
+
+// TestFilterOutStunMessages_IgnoresBytesBeyondPacket guards against classifying on
+// buffer contents left over from an earlier, longer packet.
+func TestFilterOutStunMessages_IgnoresBytesBeyondPacket(t *testing.T) {
+	buf := make([]byte, testBufSize)
+	copy(buf[4:8], magicCookieBytes)
+	buffers := [][]byte{buf}
+	bind := &ICEBind{}
+
+	filtered, err := bind.filterOutStunMessages(buffers, 2, &net.UDPAddr{})
+	assert.NoError(t, err)
+	assert.False(t, filtered, "a 2 byte packet must not be classified from stale buffer bytes")
+}
+
+// TestReceiveFn_ClearsSizeOfConsumedPacket covers the accounting WireGuard relies
+// on: sizes is reused across reads, so a slot whose packet was consumed as STUN must
+// be reported as empty. Otherwise WireGuard reprocesses the same buffer under the
+// previous packet's length, which for a WireGuard-shaped packet means it is handled
+// twice.
+func TestReceiveFn_ClearsSizeOfConsumedPacket(t *testing.T) {
+	conn := listenUDP(t, "udp4", "127.0.0.1:0")
+	defer conn.Close()
+
+	recvFn := receiverCreator{setupICEBind(t)}.CreateReceiverFn(
+		ipv4.NewPacketConn(conn), conn, false, createMsgPool(),
+	)
+
+	msg, err := stun.Build(stun.BindingRequest, stun.TransactionID, stun.Fingerprint)
+	require.NoError(t, err)
+
+	sender := listenUDP(t, "udp4", "127.0.0.1:0")
+	defer sender.Close()
+	_, err = sender.WriteTo(msg.Raw, conn.LocalAddr())
+	require.NoError(t, err)
+
+	require.NoError(t, conn.SetReadDeadline(time.Now().Add(3*time.Second)))
+
+	bufs := [][]byte{make([]byte, 1500)}
+	// A leftover size from an earlier read, which is what makes the missing reset
+	// observable.
+	sizes := []int{148}
+	eps := make([]wgConn.Endpoint, 1)
+
+	n, err := recvFn(bufs, sizes, eps)
+	require.NoError(t, err)
+	require.Equal(t, 1, n)
+	assert.Zero(t, sizes[0], "consumed STUN packet must not leave a size behind for WireGuard")
+}
+
+func TestIsTransportPkg(t *testing.T) {
+	tests := []struct {
+		name string
+		pkt  []byte
+		n    int
+		want bool
+	}{
+		{"transport data with payload", wgMsg(wgMsgTypeTransport, 128, nil), 128, true},
+		{"keepalive", wgMsg(wgMsgTypeTransport, wgMinMsgSize, nil), wgMinMsgSize, false},
+		{"handshake initiation", wgMsg(wgMsgTypeHandshakeInitiation, 148, nil), 148, false},
+		{"stale type bytes beyond packet", wgMsg(wgMsgTypeTransport, 128, nil), 2, false},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			assert.Equal(t, tc.want, isTransportPkg(intoBuffer(tc.pkt), tc.n),
+				"wrong activity classification for %s", tc.name)
+		})
+	}
+}
+
+// TestFilterOutStunMessages_ConsumesSTUNWithWireGuardShapedType covers the one STUN
+// encoding whose leading bytes collide with a WireGuard message type: method 0x080 as a
+// request encodes to 0x0200, so the type byte reads as a handshake response and the byte
+// after it is zero. Only the length check keeps such a message out of WireGuard's hands.
+// pion implements no method in that range, so this is a synthetic worst case rather than
+// traffic ICE produces.
+func TestFilterOutStunMessages_ConsumesSTUNWithWireGuardShapedType(t *testing.T) {
+	msg, err := stun.Build(stun.NewType(stun.Method(0x080), stun.ClassRequest), stun.TransactionID)
+	require.NoError(t, err)
+	require.Equal(t, []byte{0x02, 0x00, 0x00, 0x00}, msg.Raw[:4],
+		"precondition: the leading bytes read as a WireGuard message type")
+
+	buffers := intoBuffer(msg.Raw)
+	bind := &ICEBind{}
+
+	filtered, err := bind.filterOutStunMessages(buffers, len(msg.Raw), &net.UDPAddr{})
+	assert.NoError(t, err)
+	assert.True(t, filtered, "STUN message must be consumed despite its WireGuard-shaped type")
+}

client/installer.nsis

@@ -22,8 +22,6 @@
 !define UI_REG_APP_PATH "Software\Microsoft\Windows\CurrentVersion\App Paths\${UI_APP_EXE}"
 !define UI_UNINSTALL_PATH "Software\Microsoft\Windows\CurrentVersion\Uninstall\${UI_APP_NAME}"
 
-!define AUTOSTART_REG_KEY "Software\Microsoft\Windows\CurrentVersion\Run"
-
 !define NETBIRD_DATA_DIR "$COMMONPROGRAMDATA\Netbird"
 
 Unicode True
@@ -228,13 +226,6 @@ WriteRegStr ${REG_ROOT} "${UNINSTALL_PATH}"  "Publisher" "${COMP_NAME}"
 
 WriteRegStr ${REG_ROOT} "${UI_REG_APP_PATH}" "" "$INSTDIR\${UI_APP_EXE}"
 
-; Autostart is owned by the UI's per-user setting (HKCU\...\Run via Wails),
-; not the installer. Drop the machine-wide entry older installers wrote so the
-; toggle is the single source of truth. HKCU is left untouched -- it may hold
-; the user's own toggle state, which must survive upgrades.
-DetailPrint "Removing installer-managed autostart registry entry if present..."
-DeleteRegValue HKLM "${AUTOSTART_REG_KEY}" "${APP_NAME}"
-
 EnVar::SetHKLM
 EnVar::AddValueEx "path" "$INSTDIR"
 
@@ -299,15 +290,6 @@ ExecWait '"$INSTDIR\${MAIN_APP_EXE}" service uninstall'
 DetailPrint "Terminating Netbird UI process..."
 ExecWait `taskkill /im ${UI_APP_EXE}.exe /f`
 
-; Remove autostart registry entries
-DetailPrint "Removing autostart registry entries if they exist..."
-; Legacy machine-wide entry written by older installers.
-DeleteRegValue HKLM "${AUTOSTART_REG_KEY}" "${APP_NAME}"
-; Per-user entry the UI toggle writes via Wails (value name is the lowercase
-; app-name slug). Uninstall removes the app, so drop it too.
-DeleteRegValue HKCU "${AUTOSTART_REG_KEY}" "${APP_NAME}"
-DeleteRegValue HKCU "${AUTOSTART_REG_KEY}" "netbird"
-
 ; Handle data deletion based on checkbox
 DetailPrint "Checking if user requested data deletion..."
 ${If} $DeleteDataEnabled == "1"

client/internal/auth/auth.go

@@ -83,6 +83,15 @@ func NewAuth(ctx context.Context, privateKey string, mgmURL *url.URL, config *pr
 	}, nil
 }
 
+// grpcClient returns the current management connection. Callers must go through it rather than
+// reading a.client: reconnect replaces that field while other goroutines are using it.
+func (a *Auth) grpcClient() *mgm.GrpcClient {
+	a.mutex.RLock()
+	defer a.mutex.RUnlock()
+
+	return a.client
+}
+
 // Close closes the management client connection
 func (a *Auth) Close() error {
 	a.mutex.Lock()
@@ -140,25 +149,20 @@ func (a *Auth) IsSSOSupported(ctx context.Context) (bool, error) {
 // This avoids creating a new connection to the management server
 func (a *Auth) GetOAuthFlow(ctx context.Context, forceDeviceAuth bool) (OAuthFlow, error) {
 	var flow OAuthFlow
-	var err error
 
-	err = a.withRetry(ctx, func(client *mgm.GrpcClient) error {
-		if forceDeviceAuth {
-			flow, err = a.getDeviceFlow(client)
-			return err
-		}
+	// the connection is owned by a and outlives this call, so a later fallback reuses it
+	newAuth := func(context.Context) (*Auth, func(), error) {
+		return a, func() {}, nil
+	}
 
-		// Try PKCE flow first
-		flow, err = a.getPKCEFlow(client)
-		if err != nil {
-			// If PKCE not supported, try Device flow
-			if s, ok := status.FromError(err); ok && (s.Code() == codes.NotFound || s.Code() == codes.Unimplemented) {
-				flow, err = a.getDeviceFlow(client)
-				return err
-			}
-			return err
+	err := a.withRetry(ctx, func(client *mgm.GrpcClient) error {
+		var err error
+		flow, err = oauthFlowWithFallback(a, client, flowOrder(forceDeviceAuth, true), "", newAuth)
+
+		if IsSSOUnavailable(err) {
+			return backoff.Permanent(err)
 		}
-		return nil
+		return err
 	})
 
 	return flow, err

client/internal/auth/device_flow.go

@@ -48,8 +48,17 @@ type DeviceAuthProviderConfig struct {
 	LoginHint string
 }
 
-// validateDeviceAuthConfig validates device authorization provider configuration
+// validateDeviceAuthConfig validates device authorization provider configuration. A missing
+// value means management does not have this flow configured, so the error wraps
+// errFlowNotConfigured and the caller can fall back to the other flow.
 func validateDeviceAuthConfig(config *DeviceAuthProviderConfig) error {
+	if err := checkDeviceAuthConfig(config); err != nil {
+		return fmt.Errorf("%w: %w", errFlowNotConfigured, err)
+	}
+	return nil
+}
+
+func checkDeviceAuthConfig(config *DeviceAuthProviderConfig) error {
 	errorMsgFormat := "invalid provider configuration received from management: %s value is empty. Contact your NetBird administrator"
 
 	if config.Audience == "" {
@@ -161,8 +170,12 @@ func (d *DeviceAuthorizationFlow) RequestAuthInfo(ctx context.Context) (AuthFlow
 		return AuthFlowInfo{}, fmt.Errorf("reading body failed with error: %v", err)
 	}
 
-	if res.StatusCode != 200 {
-		return AuthFlowInfo{}, fmt.Errorf("request device code returned status %d error: %s", res.StatusCode, string(body))
+	if res.StatusCode != http.StatusOK {
+		reqErr := fmt.Errorf("request device code returned status %d error: %s", res.StatusCode, string(body))
+		if deviceGrantUnsupported(res.StatusCode, body) {
+			return AuthFlowInfo{}, fmt.Errorf("%w: %w", errFlowNotConfigured, reqErr)
+		}
+		return AuthFlowInfo{}, reqErr
 	}
 
 	deviceCode := AuthFlowInfo{}
@@ -186,6 +199,34 @@ func (d *DeviceAuthorizationFlow) RequestAuthInfo(ctx context.Context) (AuthFlow
 	return deviceCode, err
 }
 
+// deviceGrantUnsupported reports whether the IdP's answer to a device code request means it does
+// not serve the device authorization grant at all, rather than a transient or request-specific
+// failure. An IdP that does not route the endpoint answers 404/405/501; one that knows the
+// endpoint but has the grant disabled for this client answers with an OAuth 2.0 error code.
+func deviceGrantUnsupported(statusCode int, body []byte) bool {
+	switch statusCode {
+	case http.StatusNotFound, http.StatusMethodNotAllowed, http.StatusNotImplemented:
+		return true
+	case http.StatusBadRequest, http.StatusUnauthorized, http.StatusForbidden:
+	default:
+		return false
+	}
+
+	var oauthErr struct {
+		Error string `json:"error"`
+	}
+	if err := json.Unmarshal(body, &oauthErr); err != nil {
+		return false
+	}
+
+	switch oauthErr.Error {
+	case "unsupported_grant_type", "unauthorized_client":
+		return true
+	default:
+		return false
+	}
+}
+
 func appendLoginHint(uri, loginHint string) string {
 	if uri == "" || loginHint == "" {
 		return uri

client/internal/auth/oauth.go

@@ -2,15 +2,19 @@ package auth
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"net/http"
+	"net/url"
 	"runtime"
+	"sync"
 
 	log "github.com/sirupsen/logrus"
 	"google.golang.org/grpc/codes"
 	gstatus "google.golang.org/grpc/status"
 
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
+	mgm "github.com/netbirdio/netbird/shared/management/client"
 )
 
 // OAuthFlow represents an interface for authorization using different OAuth 2.0 flows
@@ -59,77 +63,327 @@ func (t TokenInfo) GetTokenToUse() string {
 	return t.AccessToken
 }
 
-func shouldUseDeviceFlow(force bool, isUnixDesktopClient bool) bool {
-	return force || (runtime.GOOS == "linux" || runtime.GOOS == "freebsd") && !isUnixDesktopClient
+// errFlowNotConfigured marks a flow this deployment does not offer: management returned no
+// configuration for it, the configuration it returned is incomplete, or the IdP refuses to serve
+// the grant. It is the only condition that makes the client try the other flow, so that a
+// transient failure keeps failing on the flow the user actually wants.
+var errFlowNotConfigured = errors.New("authorization flow is not configured")
+
+// ssoUnavailableError reports that the management server offers no usable SSO flow at all.
+// Retrying cannot help, so callers should surface it to the user instead of backing off.
+type ssoUnavailableError struct {
+	msg string
 }
 
-// NewOAuthFlow initializes and returns the appropriate OAuth flow based on the management configuration
-//
-// It starts by initializing the PKCE.If this process fails, it resorts to the Device Code Flow,
-// and if that also fails, the authentication process is deemed unsuccessful
-//
-// On Linux distros without desktop environment support, it only tries to initialize the Device Code Flow
-// forceDeviceCodeFlow can be used to skip PKCE and go directly to Device Code Flow (e.g., for Android TV)
-func NewOAuthFlow(ctx context.Context, config *profilemanager.Config, isUnixDesktopClient bool, forceDeviceCodeFlow bool, hint string) (OAuthFlow, error) {
-	if shouldUseDeviceFlow(forceDeviceCodeFlow, isUnixDesktopClient) {
-		return authenticateWithDeviceCodeFlow(ctx, config, hint)
-	}
-
-	pkceFlow, err := authenticateWithPKCEFlow(ctx, config, hint)
-	if err != nil {
-		log.Debugf("failed to initialize pkce authentication with error: %v\n", err)
-		log.Debug("falling back to device code flow")
-		return authenticateWithDeviceCodeFlow(ctx, config, hint)
-	}
-	return pkceFlow, nil
+func (e *ssoUnavailableError) Error() string {
+	return e.msg
 }
 
-// authenticateWithPKCEFlow initializes the Proof Key for Code Exchange flow auth flow
-func authenticateWithPKCEFlow(ctx context.Context, config *profilemanager.Config, hint string) (OAuthFlow, error) {
-	authClient, err := NewAuth(ctx, config.PrivateKey, config.ManagementURL, config)
-	if err != nil {
-		return nil, fmt.Errorf("failed to create auth client: %v", err)
-	}
-	defer authClient.Close()
+// oauthFlowInit names one of the OAuth flows and builds it from the management configuration.
+type oauthFlowInit struct {
+	name string
+	init func(a *Auth, client *mgm.GrpcClient, hint string) (OAuthFlow, error)
+}
 
-	pkceFlowInfo, err := authClient.getPKCEFlow(authClient.client)
+// authFactory hands out a management connection to build a flow with, plus the cleanup that
+// releases it. Callers that own a long-lived connection return it with a no-op cleanup.
+type authFactory func(ctx context.Context) (*Auth, func(), error)
+
+// loginHintSetter is implemented by both concrete flows but is deliberately not part of
+// OAuthFlow, so callers reach it through a type assertion.
+type loginHintSetter interface {
+	SetLoginHint(hint string)
+}
+
+// fallbackFlow wraps the flow that was picked at initialization time with the flows that were
+// not tried. Whether the IdP actually serves a flow only shows up when the flow is run: an IdP
+// with the device grant disabled answers the device code request with 404 even though
+// management handed out a device flow configuration. When that happens the wrapper swaps in the
+// next flow instead of failing the login.
+type fallbackFlow struct {
+	mu        sync.Mutex
+	active    OAuthFlow
+	remaining []oauthFlowInit
+	hint      string
+	newAuth   authFactory
+}
+
+func (f *fallbackFlow) RequestAuthInfo(ctx context.Context) (AuthFlowInfo, error) {
+	info, err := f.current().RequestAuthInfo(ctx)
+	if err == nil || !isFlowUnavailable(err) {
+		return info, err
+	}
+
+	next, nextErr := f.initNext(ctx)
+	if nextErr != nil {
+		log.Debugf("failed to fall back to another authorization flow: %v", nextErr)
+		return AuthFlowInfo{}, err
+	}
+
+	return next.RequestAuthInfo(ctx)
+}
+
+func (f *fallbackFlow) WaitToken(ctx context.Context, info AuthFlowInfo) (TokenInfo, error) {
+	return f.current().WaitToken(ctx, info)
+}
+
+func (f *fallbackFlow) GetClientID(ctx context.Context) string {
+	return f.current().GetClientID(ctx)
+}
+
+// SetLoginHint forwards the hint to the active flow and keeps it for a flow a later fallback
+// initializes. Callers that set the hint after building the flow reach the concrete flow through
+// a type assertion, which the OAuthFlow interface does not carry, so the wrapper has to offer it
+// too or the hint is silently dropped.
+func (f *fallbackFlow) SetLoginHint(hint string) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	f.hint = hint
+	if setter, ok := f.active.(loginHintSetter); ok {
+		setter.SetLoginHint(hint)
+	}
+}
+
+func (f *fallbackFlow) current() OAuthFlow {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	return f.active
+}
+
+// initNext initializes the next flow this deployment offers and makes it the active one.
+func (f *fallbackFlow) initNext(ctx context.Context) (OAuthFlow, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	if len(f.remaining) == 0 {
+		return nil, errors.New("no authorization flow left to try")
+	}
+
+	a, cleanup, err := f.newAuth(ctx)
 	if err != nil {
-		return nil, fmt.Errorf("getting pkce authorization flow info failed with error: %v", err)
+		return nil, err
+	}
+	defer cleanup()
+
+	flow, remaining, err := initFirstAvailableFlow(a, a.grpcClient(), f.remaining, f.hint)
+	if err != nil {
+		return nil, err
+	}
+
+	log.Infof("the identity provider does not serve the selected authorization flow, continuing with the next one")
+	f.active = flow
+	f.remaining = remaining
+
+	return flow, nil
+}
+
+// preferDeviceFlow reports whether the device code flow should be tried before PKCE. PKCE needs
+// a browser on this host and a loopback listener to receive the redirect, neither of which
+// exists on a Unix host without a graphical session. The GOOS guard keeps a caller that reports
+// no graphical session on a platform that always has one from changing the preference.
+func preferDeviceFlow(hasGraphicalSession bool) bool {
+	return (runtime.GOOS == "linux" || runtime.GOOS == "freebsd") && !hasGraphicalSession
+}
+
+// flowOrder returns the flows to attempt, in order.
+//
+// force leaves the device code flow on its own rather than first: it marks a device with no
+// browser at all, such as Android TV or tvOS. PKCE cannot work there even from another device,
+// because the redirect has to arrive on the loopback listener of the device being enrolled, so
+// offering it as a fallback would only replace a clear error with a login that cannot complete.
+func flowOrder(force bool, hasGraphicalSession bool) []oauthFlowInit {
+	pkce := oauthFlowInit{name: "pkce authorization flow", init: initPKCEFlow}
+	device := oauthFlowInit{name: "device code flow", init: initDeviceFlow}
+
+	switch {
+	case force:
+		return []oauthFlowInit{device}
+	case preferDeviceFlow(hasGraphicalSession):
+		return []oauthFlowInit{device, pkce}
+	default:
+		return []oauthFlowInit{pkce, device}
+	}
+}
+
+func initPKCEFlow(a *Auth, client *mgm.GrpcClient, hint string) (OAuthFlow, error) {
+	flow, err := a.getPKCEFlow(client)
+	if err != nil {
+		return nil, err
 	}
 
 	if hint != "" {
-		pkceFlowInfo.SetLoginHint(hint)
+		flow.SetLoginHint(hint)
 	}
 
-	return pkceFlowInfo, nil
+	return flow, nil
 }
 
-// authenticateWithDeviceCodeFlow initializes the Device Code auth Flow
-func authenticateWithDeviceCodeFlow(ctx context.Context, config *profilemanager.Config, hint string) (OAuthFlow, error) {
+func initDeviceFlow(a *Auth, client *mgm.GrpcClient, hint string) (OAuthFlow, error) {
+	flow, err := a.getDeviceFlow(client)
+	if err != nil {
+		return nil, err
+	}
+
+	if hint != "" {
+		flow.SetLoginHint(hint)
+	}
+
+	return flow, nil
+}
+
+// NewOAuthFlow initializes and returns an OAuth flow based on the management configuration.
+//
+// Both flows are optional server side: management answers NotFound for a flow it has no
+// configuration for. The preferred flow is tried first and the other one is used as a fallback,
+// so a server that only offers one of them still works. forceDeviceCodeFlow restricts the client
+// to the device code flow with no fallback, for a device that has no browser at all.
+func NewOAuthFlow(ctx context.Context, config *profilemanager.Config, hasGraphicalSession bool, forceDeviceCodeFlow bool, hint string) (OAuthFlow, error) {
 	authClient, err := NewAuth(ctx, config.PrivateKey, config.ManagementURL, config)
 	if err != nil {
-		return nil, fmt.Errorf("failed to create auth client: %v", err)
+		return nil, fmt.Errorf("create auth client: %w", err)
 	}
-	defer authClient.Close()
+	defer func() {
+		if err := authClient.Close(); err != nil {
+			log.Debugf("failed to close auth client: %v", err)
+		}
+	}()
 
-	deviceFlowInfo, err := authClient.getDeviceFlow(authClient.client)
+	// the connection above is closed on return, so a later fallback opens its own
+	newAuth := func(ctx context.Context) (*Auth, func(), error) {
+		a, err := NewAuth(ctx, config.PrivateKey, config.ManagementURL, config)
+		if err != nil {
+			return nil, nil, fmt.Errorf("create auth client: %w", err)
+		}
+		return a, func() {
+			if err := a.Close(); err != nil {
+				log.Debugf("failed to close auth client: %v", err)
+			}
+		}, nil
+	}
+
+	flows := flowOrder(forceDeviceCodeFlow, hasGraphicalSession)
+	return oauthFlowWithFallback(authClient, authClient.grpcClient(), flows, hint, newAuth)
+}
+
+// oauthFlowWithFallback initializes the first flow this deployment offers, moving on to the next
+// one when a flow is not configured here. It only fails once every flow has been tried, and any
+// flow left untried is handed to the returned flow so it can still fall back if the IdP rejects
+// the flow that was picked.
+func oauthFlowWithFallback(a *Auth, client *mgm.GrpcClient, flows []oauthFlowInit, hint string, newAuth authFactory) (OAuthFlow, error) {
+	flow, remaining, err := initFirstAvailableFlow(a, client, flows, hint)
 	if err != nil {
-		switch s, ok := gstatus.FromError(err); {
-		case ok && s.Code() == codes.NotFound:
-			return nil, fmt.Errorf("no SSO provider returned from management. " +
-				"Please proceed with setting up this device using setup keys " +
-				"https://docs.netbird.io/how-to/register-machines-using-setup-keys")
-		case ok && s.Code() == codes.Unimplemented:
-			return nil, fmt.Errorf("the management server, %s, does not support SSO providers, "+
-				"please update your server or use Setup Keys to login", config.ManagementURL)
-		default:
-			return nil, fmt.Errorf("getting device authorization flow info failed with error: %v", err)
+		return nil, err
+	}
+
+	if len(remaining) == 0 {
+		return flow, nil
+	}
+
+	return &fallbackFlow{
+		active:    flow,
+		remaining: remaining,
+		hint:      hint,
+		newAuth:   newAuth,
+	}, nil
+}
+
+// initFirstAvailableFlow returns the first flow that could be initialized along with the flows
+// after it, which are still untried.
+func initFirstAvailableFlow(a *Auth, client *mgm.GrpcClient, flows []oauthFlowInit, hint string) (OAuthFlow, []oauthFlowInit, error) {
+	var errs []error
+	for i, f := range flows {
+		flow, err := f.init(a, client, hint)
+		if err == nil {
+			return flow, flows[i+1:], nil
+		}
+
+		errs = append(errs, fmt.Errorf("%s: %w", f.name, err))
+
+		// only a flow this deployment does not offer is worth replacing with another one
+		if !isFlowUnavailable(err) {
+			break
+		}
+		if i < len(flows)-1 {
+			log.Infof("%s is not configured (%v), falling back to %s", f.name, err, flows[i+1].name)
 		}
 	}
 
-	if hint != "" {
-		deviceFlowInfo.SetLoginHint(hint)
+	return nil, nil, flowInitError(a.mgmURL, errs)
+}
+
+// flowInitError turns the per-flow initialization errors into a single actionable error. The
+// message stays neutral about what to do instead: SSO is also how a peer extends its session and
+// authenticates SSH, where a setup key is no alternative. Callers that are enrolling a device add
+// that advice themselves, see IsSSOUnavailable.
+func flowInitError(mgmURL *url.URL, errs []error) error {
+	if allMatch(errs, isFlowUnimplemented) {
+		return &ssoUnavailableError{msg: fmt.Sprintf("the management server, %s, does not support SSO providers, "+
+			"please update your server", mgmURL)}
 	}
 
-	return deviceFlowInfo, nil
+	if allMatch(errs, isFlowUnavailable) {
+		return &ssoUnavailableError{msg: "the management server has no SSO provider configured: " +
+			"neither the pkce authorization flow nor the device code flow is available"}
+	}
+
+	return fmt.Errorf("initialize authorization flow: %w", errors.Join(errs...))
+}
+
+// IsSSOUnavailable reports whether err means the management server offers no usable SSO flow, so
+// no retry and no other flow can help. Enrollment paths use it to point the user at setup keys.
+func IsSSOUnavailable(err error) bool {
+	var ssoUnavailable *ssoUnavailableError
+	return errors.As(err, &ssoUnavailable)
+}
+
+// WithSetupKeyAdvice appends enrollment guidance to an SSO-unavailable error and returns any
+// other error unchanged. Only enrollment can fall back to a setup key: extending a session and
+// authenticating SSH cannot, so those paths must not call this.
+//
+// The login paths that do call it cannot tell an unregistered peer from an SSO-enrolled one
+// whose session expired, since both answer PermissionDenied, so the advice names the case it
+// applies to rather than telling an enrolled peer to do something that cannot work.
+func WithSetupKeyAdvice(err error) error {
+	if !IsSSOUnavailable(err) {
+		return err
+	}
+
+	return fmt.Errorf("%w. If this device is not enrolled yet, enroll it with a setup key instead: "+
+		"https://docs.netbird.io/how-to/register-machines-using-setup-keys", err)
+}
+
+func allMatch(errs []error, match func(error) bool) bool {
+	if len(errs) == 0 {
+		return false
+	}
+
+	for _, err := range errs {
+		if !match(err) {
+			return false
+		}
+	}
+	return true
+}
+
+// isFlowUnavailable reports whether the flow is not on offer here: management has no
+// configuration for it (NotFound), predates the RPC entirely (Unimplemented), returned an
+// incomplete configuration, or the IdP does not serve the grant.
+func isFlowUnavailable(err error) bool {
+	return errors.Is(err, errFlowNotConfigured) ||
+		hasStatusCode(err, codes.NotFound) ||
+		hasStatusCode(err, codes.Unimplemented)
+}
+
+func isFlowUnimplemented(err error) bool {
+	return hasStatusCode(err, codes.Unimplemented)
+}
+
+func hasStatusCode(err error, code codes.Code) bool {
+	s, ok := gstatus.FromError(err)
+	if !ok {
+		return false
+	}
+	return s.Code() == code
 }

client/internal/auth/oauth_test.go

@@ -0,0 +1,330 @@
+package auth
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net/url"
+	"runtime"
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/status"
+
+	mgm "github.com/netbirdio/netbird/shared/management/client"
+)
+
+// stubFlow is a minimal OAuthFlow returned by the fake initializers below. requestErr, when set,
+// is what its RequestAuthInfo returns, standing in for an IdP that rejects the flow.
+type stubFlow struct {
+	name       string
+	hint       string
+	requestErr error
+}
+
+func (s *stubFlow) RequestAuthInfo(context.Context) (AuthFlowInfo, error) {
+	if s.requestErr != nil {
+		return AuthFlowInfo{}, s.requestErr
+	}
+	return AuthFlowInfo{UserCode: s.name}, nil
+}
+
+func (s *stubFlow) WaitToken(context.Context, AuthFlowInfo) (TokenInfo, error) {
+	return TokenInfo{}, nil
+}
+
+func (s *stubFlow) GetClientID(context.Context) string {
+	return ""
+}
+
+func (s *stubFlow) SetLoginHint(hint string) {
+	s.hint = hint
+}
+
+// stubInit returns a flow initializer that yields a named stub flow, or err when err is non-nil.
+func stubInit(name string, err error) oauthFlowInit {
+	return stubInitFlow(name, err, nil)
+}
+
+// stubInitFlow is stubInit with control over what the resulting flow's RequestAuthInfo returns.
+func stubInitFlow(name string, initErr, requestErr error) oauthFlowInit {
+	return oauthFlowInit{
+		name: name,
+		init: func(_ *Auth, _ *mgm.GrpcClient, hint string) (OAuthFlow, error) {
+			if initErr != nil {
+				return nil, initErr
+			}
+			return &stubFlow{name: name, hint: hint, requestErr: requestErr}, nil
+		},
+	}
+}
+
+// stubAuthFactory hands out an Auth without a management connection, which the stub
+// initializers above never touch.
+func stubAuthFactory(a *Auth) authFactory {
+	return func(context.Context) (*Auth, func(), error) {
+		return a, func() {}, nil
+	}
+}
+
+func TestOAuthFlowWithFallback(t *testing.T) {
+	notFound := status.Error(codes.NotFound, "no device authorization flow information available")
+	unimplemented := status.Error(codes.Unimplemented, "unknown method")
+	incompleteConfig := fmt.Errorf("%w: Client ID value is empty", errFlowNotConfigured)
+	unreachable := status.Error(codes.Unavailable, "connection refused")
+
+	tests := []struct {
+		name          string
+		flows         []oauthFlowInit
+		expectedFlow  string
+		expectedErr   string
+		expectedNoSSO bool
+	}{
+		{
+			name:         "preferred flow is used",
+			flows:        []oauthFlowInit{stubInit("device", nil), stubInit("pkce", nil)},
+			expectedFlow: "device",
+		},
+		{
+			// the RedHat case: device code flow disabled on management, PKCE configured
+			name:         "falls back when preferred flow is not configured",
+			flows:        []oauthFlowInit{stubInit("device", notFound), stubInit("pkce", nil)},
+			expectedFlow: "pkce",
+		},
+		{
+			name:         "falls back on an incomplete flow configuration",
+			flows:        []oauthFlowInit{stubInit("pkce", incompleteConfig), stubInit("device", nil)},
+			expectedFlow: "device",
+		},
+		{
+			name:        "does not fall back when the preferred flow fails for another reason",
+			flows:       []oauthFlowInit{stubInit("pkce", unreachable), stubInit("device", nil)},
+			expectedErr: "connection refused",
+		},
+		{
+			// stays neutral about the remedy: --extend and SSH auth cannot use a setup key
+			name:          "neither flow configured reports no SSO provider",
+			flows:         []oauthFlowInit{stubInit("device", notFound), stubInit("pkce", notFound)},
+			expectedErr:   "no SSO provider configured",
+			expectedNoSSO: true,
+		},
+		{
+			name:          "old server without the flow RPCs asks for an update",
+			flows:         []oauthFlowInit{stubInit("device", unimplemented), stubInit("pkce", unimplemented)},
+			expectedErr:   "does not support SSO providers",
+			expectedNoSSO: true,
+		},
+	}
+
+	mgmURL, err := url.Parse("https://api.netbird.io:443")
+	require.NoError(t, err)
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			a := &Auth{mgmURL: mgmURL}
+			flow, err := oauthFlowWithFallback(a, nil, tt.flows, "user@example.com", stubAuthFactory(a))
+
+			if tt.expectedErr != "" {
+				require.Error(t, err)
+				assert.Contains(t, err.Error(), tt.expectedErr)
+
+				var ssoUnavailable *ssoUnavailableError
+				assert.Equal(t, tt.expectedNoSSO, errors.As(err, &ssoUnavailable),
+					"terminal SSO-unavailable classification mismatch for %v", err)
+				return
+			}
+
+			require.NoError(t, err)
+			stub := activeStub(t, flow)
+			assert.Equal(t, tt.expectedFlow, stub.name)
+			assert.Equal(t, "user@example.com", stub.hint, "login hint must be passed to the flow")
+		})
+	}
+}
+
+// activeStub unwraps the flow currently in use, which is behind a fallbackFlow whenever an
+// untried flow is left.
+func activeStub(t *testing.T, flow OAuthFlow) *stubFlow {
+	t.Helper()
+
+	if fallback, ok := flow.(*fallbackFlow); ok {
+		flow = fallback.current()
+	}
+
+	stub, ok := flow.(*stubFlow)
+	require.True(t, ok, "unexpected flow type %T", flow)
+	return stub
+}
+
+// TestFallbackFlowRequestAuthInfo covers the failure the RedHat report hit: management hands out
+// a device flow configuration, but the IdP does not serve the grant and only says so when the
+// device code is requested.
+func TestFallbackFlowRequestAuthInfo(t *testing.T) {
+	mgmURL, err := url.Parse("https://api.netbird.io:443")
+	require.NoError(t, err)
+	a := &Auth{mgmURL: mgmURL}
+
+	idpRejects := fmt.Errorf("%w: request device code returned status 404", errFlowNotConfigured)
+
+	t.Run("swaps in the untried flow", func(t *testing.T) {
+		flows := []oauthFlowInit{stubInitFlow("device", nil, idpRejects), stubInit("pkce", nil)}
+
+		flow, err := oauthFlowWithFallback(a, nil, flows, "", stubAuthFactory(a))
+		require.NoError(t, err)
+		require.Equal(t, "device", activeStub(t, flow).name)
+
+		info, err := flow.RequestAuthInfo(context.Background())
+		require.NoError(t, err)
+		assert.Equal(t, "pkce", info.UserCode, "the request must be served by the fallback flow")
+		assert.Equal(t, "pkce", activeStub(t, flow).name, "the fallback flow must stay active for WaitToken")
+	})
+
+	t.Run("keeps the original error when nothing else is configured", func(t *testing.T) {
+		flows := []oauthFlowInit{
+			stubInitFlow("device", nil, idpRejects),
+			stubInit("pkce", status.Error(codes.NotFound, "no pkce authorization flow information available")),
+		}
+
+		flow, err := oauthFlowWithFallback(a, nil, flows, "", stubAuthFactory(a))
+		require.NoError(t, err)
+
+		_, err = flow.RequestAuthInfo(context.Background())
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), "status 404")
+	})
+
+	t.Run("keeps the original error when the fallback cannot reach management", func(t *testing.T) {
+		flows := []oauthFlowInit{stubInitFlow("device", nil, idpRejects), stubInit("pkce", nil)}
+
+		unreachable := func(context.Context) (*Auth, func(), error) {
+			return nil, nil, errors.New("connect to management: connection refused")
+		}
+
+		flow, err := oauthFlowWithFallback(a, nil, flows, "", unreachable)
+		require.NoError(t, err)
+
+		_, err = flow.RequestAuthInfo(context.Background())
+		require.Error(t, err)
+		assert.Contains(t, err.Error(), "status 404", "the IdP error must survive a failed fallback")
+		assert.Equal(t, "device", activeStub(t, flow).name, "a failed fallback must not swap the flow")
+	})
+
+	t.Run("does not swap flows on an unrelated failure", func(t *testing.T) {
+		flows := []oauthFlowInit{
+			stubInitFlow("device", nil, errors.New("timeout talking to the IdP")),
+			stubInit("pkce", nil),
+		}
+
+		flow, err := oauthFlowWithFallback(a, nil, flows, "", stubAuthFactory(a))
+		require.NoError(t, err)
+
+		_, err = flow.RequestAuthInfo(context.Background())
+		require.Error(t, err)
+		assert.Equal(t, "device", activeStub(t, flow).name, "the preferred flow must stay active")
+	})
+}
+
+// TestForcedDeviceFlowHasNoFallback covers Android TV and tvOS: a browserless device must get the
+// device code error rather than a PKCE flow it can never complete.
+func TestForcedDeviceFlowHasNoFallback(t *testing.T) {
+	mgmURL, err := url.Parse("https://api.netbird.io:443")
+	require.NoError(t, err)
+	a := &Auth{mgmURL: mgmURL}
+
+	notFound := status.Error(codes.NotFound, "no device authorization flow information available")
+
+	t.Run("no wrapper when the device flow works", func(t *testing.T) {
+		// flowOrder(force) yields this single-entry list, see TestFlowOrder
+		forced := []oauthFlowInit{stubInit("device", nil)}
+
+		flow, err := oauthFlowWithFallback(a, nil, forced, "", stubAuthFactory(a))
+		require.NoError(t, err)
+
+		_, wrapped := flow.(*fallbackFlow)
+		assert.False(t, wrapped, "nothing may swap the flow later on a browserless device")
+	})
+
+	t.Run("reports the device flow error instead of falling back", func(t *testing.T) {
+		forced := []oauthFlowInit{stubInit("device", notFound)}
+
+		_, err := oauthFlowWithFallback(a, nil, forced, "", stubAuthFactory(a))
+		require.Error(t, err)
+		assert.True(t, IsSSOUnavailable(err), "the caller must see that SSO is unavailable here")
+	})
+}
+
+// TestFallbackFlowSetLoginHint covers the Android SDK's pattern: it sets the login hint after the
+// flow is built, through a type assertion that the wrapper must satisfy.
+func TestFallbackFlowSetLoginHint(t *testing.T) {
+	mgmURL, err := url.Parse("https://api.netbird.io:443")
+	require.NoError(t, err)
+	a := &Auth{mgmURL: mgmURL}
+
+	idpRejects := fmt.Errorf("%w: request device code returned status 404", errFlowNotConfigured)
+	flows := []oauthFlowInit{stubInitFlow("device", nil, idpRejects), stubInit("pkce", nil)}
+
+	flow, err := oauthFlowWithFallback(a, nil, flows, "", stubAuthFactory(a))
+	require.NoError(t, err)
+
+	setter, ok := flow.(loginHintSetter)
+	require.True(t, ok, "the wrapper must accept a login hint like the concrete flows do")
+	setter.SetLoginHint("user@example.com")
+	assert.Equal(t, "user@example.com", activeStub(t, flow).hint, "the active flow must get the hint")
+
+	// the device flow is rejected by the IdP here, so the hint has to survive into the fallback
+	_, err = flow.RequestAuthInfo(context.Background())
+	require.NoError(t, err)
+	assert.Equal(t, "pkce", activeStub(t, flow).name)
+	assert.Equal(t, "user@example.com", activeStub(t, flow).hint, "the fallback flow must get the hint too")
+}
+
+func TestWithSetupKeyAdvice(t *testing.T) {
+	other := errors.New("connection refused")
+	assert.Equal(t, other, WithSetupKeyAdvice(other), "only an SSO-unavailable error gets advice")
+
+	advised := WithSetupKeyAdvice(&ssoUnavailableError{msg: "no SSO provider configured"})
+	assert.Contains(t, advised.Error(), "no SSO provider configured", "the original message must survive")
+	assert.Contains(t, advised.Error(), "setup key")
+	// a setup key cannot re-enrol a peer whose SSO session expired, and the login paths cannot
+	// tell that peer apart from an unregistered one, so the advice must state its condition
+	assert.Contains(t, advised.Error(), "not enrolled yet")
+	assert.True(t, IsSSOUnavailable(advised), "advice must keep the error classifiable")
+}
+
+func flowNames(flows []oauthFlowInit) []string {
+	names := make([]string, 0, len(flows))
+	for _, f := range flows {
+		names = append(names, f.name)
+	}
+	return names
+}
+
+func TestFlowOrder(t *testing.T) {
+	const pkce, device = "pkce authorization flow", "device code flow"
+
+	assert.Equal(t, []string{pkce, device}, flowNames(flowOrder(false, true)),
+		"a device with a browser tries PKCE first and keeps the device code flow as a fallback")
+
+	// only a unix host without a graphical session lacks a browser; the other platforms have one
+	headless := []string{pkce, device}
+	if runtime.GOOS == "linux" || runtime.GOOS == "freebsd" {
+		headless = []string{device, pkce}
+	}
+	assert.Equal(t, headless, flowNames(flowOrder(false, false)), "on %s", runtime.GOOS)
+
+	// Android TV and tvOS have no browser, so PKCE cannot complete there even from another
+	// device: the redirect must reach the loopback listener of the device being enrolled.
+	assert.Equal(t, []string{device}, flowNames(flowOrder(true, false)),
+		"a forced device code flow must not fall back to PKCE")
+	assert.Equal(t, []string{device}, flowNames(flowOrder(true, true)),
+		"force wins over a reported graphical session")
+}
+
+func TestPreferDeviceFlow(t *testing.T) {
+	isUnix := runtime.GOOS == "linux" || runtime.GOOS == "freebsd"
+
+	assert.Equal(t, isUnix, preferDeviceFlow(false), "headless unix hosts prefer the device flow")
+	assert.False(t, preferDeviceFlow(true), "clients with a graphical session prefer PKCE")
+}

client/internal/auth/pkce_flow.go

@@ -62,8 +62,17 @@ type PKCEAuthProviderConfig struct {
 	LoginHint string
 }
 
-// validatePKCEConfig validates PKCE provider configuration
+// validatePKCEConfig validates PKCE provider configuration. A missing value means management
+// does not have this flow configured, so the error wraps errFlowNotConfigured and the caller can
+// fall back to the other flow.
 func validatePKCEConfig(config *PKCEAuthProviderConfig) error {
+	if err := checkPKCEConfig(config); err != nil {
+		return fmt.Errorf("%w: %w", errFlowNotConfigured, err)
+	}
+	return nil
+}
+
+func checkPKCEConfig(config *PKCEAuthProviderConfig) error {
 	errorMsgFormat := "invalid provider configuration received from management: %s value is empty. Contact your NetBird administrator"
 
 	if config.ClientID == "" {

client/internal/debug/debug.go

@@ -34,9 +34,8 @@ import (
 	"github.com/netbirdio/netbird/shared/netiputil"
 )
 
-const readmeContent = `Netbird debug bundle
-This debug bundle contains the following files.
-If the --anonymize flag is set, the files are anonymized to protect sensitive information.
+const readmeContent = `This debug bundle contains the following files.
+If anonymization is enabled (--anonymize / --anonymize-level), the files are anonymized to protect sensitive information.
 
 status.txt: Anonymized status information of the NetBird client.
 client.log: Most recent, anonymized client log file of the NetBird client.
@@ -70,21 +69,34 @@ capture.pcap: Packet capture in pcap format. Only present when capture was runni
 
 
 Anonymization Process
-The files in this bundle have been anonymized to protect sensitive information. Here's how the anonymization was applied:
+The files in this bundle have been anonymized to protect sensitive information. The level applied to this bundle is recorded at the top of this file. Here's how the anonymization was applied:
 
 IP Addresses
 
-IPv4 addresses are replaced with addresses starting from 198.51.100.0
-IPv6 addresses are replaced with addresses starting from 100::
+Default level:
+- Public IPv4 addresses are replaced with addresses starting from 198.51.100.0
+- Public IPv6 addresses are replaced with addresses starting from 2001:db8:ffff::
+- IPv6 unique local addresses (fc00::/7) are anonymized as well: their random global ID uniquely identifies the network.
+- IP addresses from internal IPv4 ranges and well-known addresses are not anonymized (e.g. 8.8.8.8, 100.64.0.0/10, addresses starting with 192.168., 172.16., 10., 169.254., fe80::).
+
+Strict level (--anonymize-level strict), in addition to the default level:
+- Private (RFC 1918), CGNAT (100.64.0.0/10), and link-local (169.254.0.0/16, fe80::/10) addresses are anonymized too.
+- Internal IPv4 addresses are replaced with addresses starting from 198.18.0.0 and internal IPv6 addresses with addresses starting from 2001:db8:1::, so internal addresses remain distinguishable from public ones.
+- Addresses are mapped in order of first appearance: subnet structure, allocation scheme, and gateway conventions are not preserved. Prefix lengths of networks are preserved.
+- Peer names in front of NetBird domains are replaced with numbered placeholders (e.g. peer-1.netbird.cloud), and subdomain labels of other domains with host-N placeholders.
+- WireGuard public keys are replaced with consistent placeholder keys.
 
-IP addresses from non public ranges and well known addresses are not anonymized (e.g. 8.8.8.8, 100.64.0.0/10, addresses starting with 192.168., 172.16., 10., etc.).
 Reoccuring IP addresses are replaced with the same anonymized address.
 
 Note: The anonymized IP addresses in the status file do not match those in the log and routes files. However, the anonymized IP addresses are consistent within the status file and across the routes and log files.
 
+MAC Addresses
+MAC addresses are replaced at every anonymization level with consistent placeholders counting up from 02:00:00:00:00:01. Broadcast, multicast, and all-zero addresses are kept. At the default level a preserved IPv6 link-local address may still embed a MAC address (EUI-64); the strict level anonymizes those addresses.
+
 Domains
 All domain names (except for the netbird domains) are replaced with randomly generated strings ending in ".domain". Anonymized domains are consistent across all files in the bundle.
 Reoccuring domain names are replaced with the same anonymized domain.
+At the strict level, the peer name labels in front of netbird domains are anonymized as well.
 
 Sync Response
 The network_map.json file contains the following anonymized information:
@@ -281,6 +293,7 @@ type BundleGenerator struct {
 	cliVersion     string
 
 	anonymize         bool
+	anonymizeLevel    anonymize.Level
 	includeSystemInfo bool
 	logFileCount      uint32
 
@@ -288,7 +301,10 @@ type BundleGenerator struct {
 }
 
 type BundleConfig struct {
-	Anonymize         bool
+	Anonymize bool
+	// AnonymizeLevel selects how much the anonymizer redacts.
+	// anonymize.LevelStrict implies Anonymize.
+	AnonymizeLevel    anonymize.Level
 	IncludeSystemInfo bool
 	LogFileCount      uint32
 }
@@ -327,8 +343,11 @@ func NewBundleGenerator(deps GeneratorDependencies, cfg BundleConfig) *BundleGen
 		uiLogOpener = openLogFile
 	}
 
+	anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+	anonymizer.SetLevel(cfg.AnonymizeLevel)
+
 	return &BundleGenerator{
-		anonymizer: anonymize.NewAnonymizer(anonymize.DefaultAddresses()),
+		anonymizer: anonymizer,
 
 		internalConfig: deps.InternalConfig,
 		statusRecorder: deps.StatusRecorder,
@@ -345,7 +364,8 @@ func NewBundleGenerator(deps GeneratorDependencies, cfg BundleConfig) *BundleGen
 		daemonVersion:  deps.DaemonVersion,
 		cliVersion:     deps.CliVersion,
 
-		anonymize:         cfg.Anonymize,
+		anonymize:         cfg.Anonymize || cfg.AnonymizeLevel >= anonymize.LevelStrict,
+		anonymizeLevel:    cfg.AnonymizeLevel,
 		includeSystemInfo: cfg.IncludeSystemInfo,
 		logFileCount:      logFileCount,
 	}
@@ -485,7 +505,13 @@ func (g *BundleGenerator) addSystemInfo() {
 }
 
 func (g *BundleGenerator) addReadme() error {
-	readmeReader := strings.NewReader(readmeContent)
+	level := "none (anonymization disabled)"
+	if g.anonymize {
+		level = g.anonymizeLevel.String()
+	}
+	header := fmt.Sprintf("Netbird debug bundle\nAnonymization level applied to this bundle: %s\n", level)
+
+	readmeReader := strings.NewReader(header + readmeContent)
 	if err := g.addFileToZip(readmeReader, "README.txt"); err != nil {
 		return fmt.Errorf("add README file to zip: %w", err)
 	}
@@ -507,9 +533,10 @@ func (g *BundleGenerator) addStatus() error {
 		fullStatus := g.statusRecorder.GetFullStatus()
 		protoFullStatus := nbstatus.ToProtoFullStatus(fullStatus)
 		overview := nbstatus.ConvertToStatusOutputOverview(protoFullStatus, nbstatus.ConvertOptions{
-			Anonymize:     g.anonymize,
-			ProfileName:   profName,
-			DaemonVersion: g.daemonVersion,
+			Anonymize:      g.anonymize,
+			AnonymizeLevel: g.anonymizeLevel,
+			ProfileName:    profName,
+			DaemonVersion:  g.daemonVersion,
 		})
 		overview.CliVersion = g.cliVersion
 		statusOutput := overview.FullDetailSummary()
@@ -662,7 +689,7 @@ func (g *BundleGenerator) addCommonConfigFields(configContent *strings.Builder)
 	configContent.WriteString("NetBird Client Configuration:\n\n")
 
 	if key, err := wgtypes.ParseKey(g.internalConfig.PrivateKey); err == nil {
-		configContent.WriteString(fmt.Sprintf("PublicKey: %s\n", key.PublicKey().String()))
+		configContent.WriteString(fmt.Sprintf("PublicKey: %s\n", g.anonymizer.AnonymizeWGKey(key.PublicKey().String())))
 	}
 	configContent.WriteString(fmt.Sprintf("WgIface: %s\n", g.internalConfig.WgIface))
 	configContent.WriteString(fmt.Sprintf("WgPort: %d\n", g.internalConfig.WgPort))
@@ -952,6 +979,11 @@ func (g *BundleGenerator) addUpdateLogs() error {
 		}
 
 		baseName := filepath.Base(logFile)
+		data, err = g.anonymizeBytes(data)
+		if err != nil {
+			log.Warnf("skipping update log file %s: %v", baseName, err)
+			continue
+		}
 		if err := g.addFileToZip(bytes.NewReader(data), filepath.Join("update-logs", baseName)); err != nil {
 			return fmt.Errorf("add update log file %s to zip: %w", baseName, err)
 		}
@@ -979,6 +1011,13 @@ func (g *BundleGenerator) addCorruptedStateFiles() error {
 		}
 
 		fileName := filepath.Base(match)
+		// Corrupted state files usually fail structured JSON anonymization,
+		// so run them through the string anonymizer instead.
+		data, err = g.anonymizeBytes(data)
+		if err != nil {
+			log.Warnf("skipping corrupted state file %s: %v", fileName, err)
+			continue
+		}
 		if err := g.addFileToZip(bytes.NewReader(data), "corrupted_states/"+fileName); err != nil {
 			log.Warnf("Failed to add corrupted state file %s to zip: %v", fileName, err)
 			continue
@@ -990,6 +1029,27 @@ func (g *BundleGenerator) addCorruptedStateFiles() error {
 	return nil
 }
 
+// anonymizeBytes runs raw file content through the string anonymizer line by
+// line when anonymization is enabled. It errors instead of returning partial
+// content, so a caller never adds an unanonymized fallback to the bundle.
+func (g *BundleGenerator) anonymizeBytes(data []byte) ([]byte, error) {
+	if !g.anonymize {
+		return data, nil
+	}
+
+	var buf bytes.Buffer
+	scanner := bufio.NewScanner(bytes.NewReader(data))
+	scanner.Buffer(make([]byte, 1024*1024), 1024*1024)
+	for scanner.Scan() {
+		buf.WriteString(g.anonymizer.AnonymizeString(scanner.Text()))
+		buf.WriteByte('\n')
+	}
+	if err := scanner.Err(); err != nil {
+		return nil, fmt.Errorf("anonymize content: %w", err)
+	}
+	return buf.Bytes(), nil
+}
+
 func (g *BundleGenerator) addMetrics() error {
 	if g.clientMetrics == nil {
 		log.Debugf("skipping metrics in debug bundle: no metrics collector")
@@ -1462,6 +1522,7 @@ func anonymizeRemotePeer(peer *mgmProto.RemotePeerConfig, anonymizer *anonymize.
 	}
 
 	peer.Fqdn = anonymizer.AnonymizeDomain(peer.Fqdn)
+	peer.WgPubKey = anonymizer.AnonymizeWGKey(peer.WgPubKey)
 
 	anonymizeSSHConfig(peer.SshConfig)
 }

client/internal/debug/debug_linux.go

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

client/internal/debug/wgshow.go

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

client/internal/dns/host_darwin.go

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

client/internal/dns/host_darwin_test.go

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

client/internal/engine.go

@@ -4,7 +4,6 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"math"
 	"math/rand"
 	"net"
 	"net/netip"
@@ -24,6 +23,7 @@ import (
 	"golang.zx2c4.com/wireguard/tun/netstack"
 	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
 
+	"github.com/netbirdio/netbird/client/anonymize"
 	nberrors "github.com/netbirdio/netbird/client/errors"
 	"github.com/netbirdio/netbird/client/firewall"
 	"github.com/netbirdio/netbird/client/firewall/firewalld"
@@ -51,7 +51,6 @@ import (
 	icemaker "github.com/netbirdio/netbird/client/internal/peer/ice"
 	"github.com/netbirdio/netbird/client/internal/peerstore"
 	"github.com/netbirdio/netbird/client/internal/portforward"
-	"github.com/netbirdio/netbird/client/internal/pqkem"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/internal/relay"
 	"github.com/netbirdio/netbird/client/internal/rosenpass"
@@ -199,10 +198,6 @@ type Engine struct {
 	// rpManager is a Rosenpass manager
 	rpManager *rosenpass.Manager
 
-	// pqkemManager runs the ML-KEM post-quantum PSK exchange (gated by NB_ENABLE_PQ_MLKEM).
-	// It owns the data-path transport and peer endpoint routing.
-	pqkemManager *pqkem.Manager
-
 	// syncMsgMux is used to guarantee sequential Management Service message processing
 	syncMsgMux *sync.Mutex
 
@@ -561,11 +556,7 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 	publicKey := e.config.WgPrivateKey.PublicKey()
 	e.flowManager = netflow.NewManager(e.wgInterface, publicKey[:], e.statusRecorder)
 
-	// Rosenpass and ML-KEM are mutually exclusive. ML-KEM (NB_ENABLE_PQ_MLKEM) takes precedence
-	if e.config.RosenpassEnabled && pqkem.Enabled() {
-		log.Warnf("rosenpass and ML-KEM post-quantum are mutually exclusive; ML-KEM is enabled, so rosenpass is disabled")
-	}
-	if e.config.RosenpassEnabled && !pqkem.Enabled() {
+	if e.config.RosenpassEnabled {
 		log.Infof("rosenpass is enabled")
 		if e.config.RosenpassPermissive {
 			log.Infof("running rosenpass in permissive mode")
@@ -654,35 +645,6 @@ func (e *Engine) Start(netbirdConfig *mgmProto.NetbirdConfig, mgmtURL *url.URL)
 		e.rpManager.SetInterface(e.wgInterface)
 	}
 
-	// Start the ML-KEM PQ manager after the interface is up so its dedicated UDP
-	// transport can bind on the WG overlay IP.
-	if pqkem.Enabled() {
-		tr, pqErr := newPQTransport(e.config.WgAddr.IP)
-		if pqErr != nil {
-			// In strict mode the peer must fail closed; silently continuing without the PQ
-			// exchange would hand out classic tunnels, so treat the bind failure as fatal.
-			if pqkem.Strict() {
-				return fmt.Errorf("pqkem: strict mode enabled but transport bind failed: %w", pqErr)
-			}
-			log.Errorf("pqkem: transport bind failed, exchange disabled: %v", pqErr)
-		} else {
-			cbHandler := pqCallbackHandler{
-				wg: e.wgInterface,
-				// On a persistent rekey failure, re-bootstrap the KEM over Signal: a
-				// fresh signalling offer starts a new exchange that overwrites the
-				// stalled PSK on both sides, recovering from a data-path desync.
-				reoffer: func(remoteKey string) {
-					if conn, ok := e.peerStore.PeerConn(remoteKey); ok {
-						conn.RequestReoffer()
-					}
-				},
-			}
-			e.pqkemManager = pqkem.NewManager(pqkem.LocalID(publicKey.String()), cbHandler, pqkem.NewLogger())
-			e.pqkemManager.Start(tr)
-			log.Infof("pqkem: enabled (udp port %d on overlay %s)", e.pqkemManager.LocalPort(), e.config.WgAddr.IP)
-		}
-	}
-
 	// if inbound conns are blocked there is no need to create the ACL manager
 	if e.firewall != nil && !e.config.BlockInbound {
 		e.acl = acl.NewDefaultManager(e.firewall)
@@ -946,10 +908,6 @@ func (e *Engine) removePeer(peerKey string) error {
 
 	e.connMgr.RemovePeerConn(peerKey)
 
-	if e.pqkemManager != nil {
-		e.pqkemManager.RemovePeer(pqkem.RemoteID(peerKey))
-	}
-
 	err := e.statusRecorder.RemovePeer(peerKey)
 	if err != nil {
 		log.Warnf("received error when removing peer %s from status recorder: %v", peerKey, err)
@@ -1428,6 +1386,7 @@ func (e *Engine) handleBundle(params *mgmProto.BundleParameters) (*mgmProto.JobR
 
 	bundleJobParams := debug.BundleConfig{
 		Anonymize:         params.Anonymize,
+		AnonymizeLevel:    anonymize.ParseLevel(params.AnonymizeLevel),
 		IncludeSystemInfo: true,
 		LogFileCount:      uint32(params.LogFileCount),
 	}
@@ -1936,10 +1895,6 @@ func (e *Engine) createPeerConn(pubKey string, allowedIPs []netip.Prefix, agentV
 		},
 		ICEConfig: e.createICEConfig(),
 	}
-	if e.pqkemManager != nil {
-		config.PQ = pqHandshaker{mgr: e.pqkemManager}
-		config.PQStrict = pqkem.Strict()
-	}
 
 	serviceDependencies := peer.ServiceDependencies{
 		StatusRecorder:     e.statusRecorder,
@@ -2123,10 +2078,6 @@ func (e *Engine) close() {
 		_ = e.rpManager.Close()
 	}
 
-	if e.pqkemManager != nil {
-		e.pqkemManager.Stop()
-	}
-
 	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
 	defer cancel()
 	if err := e.portForwardManager.GracefullyStop(ctx); err != nil {
@@ -2894,13 +2845,6 @@ func convertToOfferAnswer(msg *sProto.Message) (*peer.OfferAnswer, error) {
 
 	relayIP := decodeRelayIP(msg.GetBody().GetRelayServerIP())
 
-	// Ports are uint16 internally; the proto widens them to uint32, so validate the
-	// range before narrowing (a value that does not fit is a malformed message).
-	mlkemPort := msg.GetBody().GetMlkemPort()
-	if mlkemPort > math.MaxUint16 {
-		return nil, fmt.Errorf("invalid ML-KEM port %d in signalling message", mlkemPort)
-	}
-
 	offerAnswer := peer.OfferAnswer{
 		IceCredentials: peer.IceCredentials{
 			UFrag: remoteCred.UFrag,
@@ -2910,8 +2854,6 @@ func convertToOfferAnswer(msg *sProto.Message) (*peer.OfferAnswer, error) {
 		Version:         msg.GetBody().GetNetBirdVersion(),
 		RosenpassPubKey: rosenpassPubKey,
 		RosenpassAddr:   rosenpassAddr,
-		MlkemPayload:    msg.GetBody().GetMlkemPayload(),
-		MlkemPort:       uint16(mlkemPort),
 		RelaySrvAddress: msg.GetBody().GetRelayServerAddress(),
 		RelaySrvIP:      relayIP,
 		SessionID:       sessionID,

client/internal/metrics/connection_type.go

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

client/internal/metrics/influxdb_test.go

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

client/internal/metrics/infra/README.md

@@ -56,14 +56,33 @@ Measurement: `netbird_peer_connection`
 
 Tags:
 - `deployment_type`: "cloud" | "selfhosted" | "unknown"
-- `connection_type`: "ice" | "relay"
+- `connection_type`: "ice_p2p" | "ice_turn" | "relay" (see below)
 - `attempt_type`: "initial" | "reconnection"
 - `version`: NetBird version string
 - `os`: Operating system (linux, darwin, windows, android, ios, etc.)
 - `arch`: CPU architecture (amd64, arm64, etc.)
+- `peer_id`: anonymised peer identifier (truncated SHA-256 of the WireGuard public key)
+- `connection_pair_id`: deterministic identifier for the peer pair, identical on both sides
 
 **Note:** `SignalingReceived` is set when the first offer or answer arrives from the remote peer (in both initial and reconnection paths). It excludes the potentially unbounded wait for the remote peer to come online.
 
+#### `connection_type` values
+
+Derived from the connection priority (`conntype.ConnPriority`) by `metricsConnType` in `client/internal/peer/conn.go`:
+
+| Value | Priority | Traffic is |
+|-------|----------|------------|
+| `ice_p2p` | `ICEP2P` | direct peer-to-peer |
+| `ice_turn` | `ICETurn` | relayed, through a TURN server |
+| `relay` | `Relay` | relayed, through a NetBird relay |
+| `unknown` | `None` or unrecognised | no active transport — **the sample is not pushed** |
+
+**Direct traffic is `ice_p2p` only.** `ice_turn` is relayed despite being negotiated by ICE, matching `Conn.isRelayed`.
+
+`None` means no transport is active: not established yet, or reset after a relay drop or a peer-state reset. Such a sample cannot be attributed to a transport, so `recordConnectionMetrics` drops it instead of pushing it — `unknown` therefore never appears in the bucket. Connection counts are counts of connections whose transport was known at sampling time.
+
+**Samples recorded before 0.77 used a single `ice` value** which covered `ICEP2P`, `ICETurn` *and* `None`, so historical `ice` samples overstate direct connections by an unknown amount and must not be compared with `ice_p2p`.
+
 ### Sync Duration
 
 Measurement: `netbird_sync`

client/internal/peer/conn.go

@@ -3,7 +3,6 @@ package peer
 import (
 	"context"
 	"fmt"
-	"math"
 	"net"
 	"net/netip"
 	"runtime"
@@ -27,7 +26,6 @@ import (
 	"github.com/netbirdio/netbird/client/internal/portforward"
 	"github.com/netbirdio/netbird/client/internal/rosenpass"
 	"github.com/netbirdio/netbird/client/internal/stdnet"
-	"github.com/netbirdio/netbird/monotime"
 	"github.com/netbirdio/netbird/route"
 	relayClient "github.com/netbirdio/netbird/shared/relay/client"
 )
@@ -76,39 +74,6 @@ type RosenpassConfig struct {
 	PermissiveMode bool
 }
 
-// PQHandshaker attaches post-quantum ML-KEM material to signalling offers/answers and
-// feeds received material back. It is implemented by the engine over the pqkem
-// manager and is nil when the PQ exchange is disabled. remoteKey is the peer's
-// WireGuard public key.
-type PQHandshaker interface {
-	// OfferPayload returns the KEM offer to embed in an outgoing offer (nil if this
-	// peer is not the KEM initiator) and the local PQ data-path port to announce.
-	OfferPayload(remoteKey string) (payload []byte, port uint16)
-	// ShouldSendBootstrapOffer reports whether, as the controller, we should reply to a
-	// received responder offer with our own KEM offer (true only when no exchange is
-	// already in flight — so we kick the KEM once and ignore further offers).
-	ShouldSendBootstrapOffer(remoteKey string) bool
-	// AnswerPayload processes a received KEM offer (nil if absent) and returns the KEM
-	// answer to embed in the outgoing answer (nil if none) and the local PQ port.
-	AnswerPayload(remoteKey string, recvOffer []byte) (payload []byte, port uint16)
-	// OnAnswer feeds a received KEM answer (nil if absent).
-	OnAnswer(remoteKey string, recvAnswer []byte)
-	// PSK returns the peer's latest derived post-quantum PSK to program at WG
-	// peer-config time (the pull path). ok is false until one has been derived.
-	PSK(remoteKey string) (wgtypes.Key, bool)
-	// SetRemoteAddr registers the peer's data-path endpoint learned from signalling:
-	// its WG overlay IP with the announced pq UDP port (port 0 means the peer omitted
-	// it and is on the default port).
-	SetRemoteAddr(remoteKey string, addr netip.AddrPort)
-	// OnDataPathRekeyed signals a fresh WireGuard handshake for the peer; it clocks the
-	// next chained PSK rotation pushed over the data path. sinceActivity is how long
-	// ago the peer last exchanged real user data, so the rotation can be skipped for
-	// idle tunnels.
-	OnDataPathRekeyed(remoteKey string, sinceActivity time.Duration)
-	// OnDataPathDown signals the peer's tunnel went down.
-	OnDataPathDown(remoteKey string)
-}
-
 // ConnConfig is a peer Connection configuration
 type ConnConfig struct {
 	// Key is a public key of a remote peer
@@ -126,12 +91,6 @@ type ConnConfig struct {
 
 	RosenpassConfig RosenpassConfig
 
-	// PQ carries post-quantum ML-KEM material on offers/answers; nil when disabled.
-	PQ PQHandshaker
-	// PQStrict fails closed: block peer traffic until the ML-KEM PSK is established,
-	// instead of letting the tunnel come up classically and upgrading to PQ later.
-	PQStrict bool
-
 	// ICEConfig ICE protocol configuration
 	ICEConfig icemaker.Config
 }
@@ -190,11 +149,6 @@ type Conn struct {
 	// pendingFirstPacket is the lazyconn-captured handshake init, replayed once the real
 	// transport is up.
 	pendingFirstPacket []byte
-
-	// pqStrictSentinelKey is a per-conn random sentinel PSK used in PQ strict mode to
-	// fail closed: it is programmed until the real ML-KEM PSK is derived, so no session
-	// can form on a non-PQ key. Per-conn random so two strict peers never match by chance.
-	pqStrictSentinelKey *wgtypes.Key
 }
 
 // injectPendingFirstPacket replays the captured handshake through the proxy if present, else
@@ -252,16 +206,6 @@ func NewConn(config ConnConfig, services ServiceDependencies) (*Conn, error) {
 		metricsRecorder:    services.MetricsRecorder,
 	}
 
-	if config.PQ != nil && config.PQStrict {
-		// The sentinel is what makes strict mode fail closed; if we cannot generate it we
-		// must not fall back to a usable key, so fail creating the conn instead.
-		k, err := wgtypes.GenerateKey()
-		if err != nil {
-			return nil, fmt.Errorf("generate pqkem strict-mode sentinel key: %w", err)
-		}
-		conn.pqStrictSentinelKey = &k
-	}
-
 	return conn, nil
 }
 
@@ -463,9 +407,6 @@ func (conn *Conn) ConnID() id.ConnID {
 
 // configureConnection starts proxying traffic from/to local Wireguard and sets connection status to StatusConnected
 func (conn *Conn) onICEConnectionIsReady(priority conntype.ConnPriority, iceConnInfo ICEConnInfo) {
-	// Read the PQ PSK before conn.mu to keep the lock order conn.mu -> manager.
-	pqPSK, pqOK := conn.pqPSK()
-
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
 
@@ -483,7 +424,7 @@ func (conn *Conn) onICEConnectionIsReady(priority conntype.ConnPriority, iceConn
 	if conn.currentConnPriority > priority {
 		conn.Log.Infof("current connection priority (%s) is higher than the new one (%s), do not upgrade connection", conn.currentConnPriority, priority)
 		conn.statusICE.SetConnected()
-		conn.updateIceState(iceConnInfo, pqOK, time.Now())
+		conn.updateIceState(iceConnInfo, time.Now())
 		return
 	}
 
@@ -526,7 +467,7 @@ func (conn *Conn) onICEConnectionIsReady(priority conntype.ConnPriority, iceConn
 	updateTime := time.Now()
 	conn.enableWgWatcherIfNeeded(updateTime)
 
-	presharedKey := conn.presharedKey(iceConnInfo.RosenpassPubKey, pqPSK)
+	presharedKey := conn.presharedKey(iceConnInfo.RosenpassPubKey)
 	if err = conn.endpointUpdater.ConfigureWGEndpoint(ep, presharedKey); err != nil {
 		conn.handleConfigurationFailure(err, wgProxy)
 		return
@@ -542,14 +483,11 @@ func (conn *Conn) onICEConnectionIsReady(priority conntype.ConnPriority, iceConn
 
 	conn.currentConnPriority = priority
 	conn.statusICE.SetConnected()
-	conn.updateIceState(iceConnInfo, pqOK, updateTime)
+	conn.updateIceState(iceConnInfo, updateTime)
 	conn.doOnConnected(iceConnInfo.RosenpassPubKey, iceConnInfo.RosenpassAddr, updateTime)
 }
 
 func (conn *Conn) onICEStateDisconnected(sessionChanged bool) {
-	// Read the PQ PSK before conn.mu to keep the lock order conn.mu -> manager.
-	pqPSK, _ := conn.pqPSK()
-
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
 
@@ -576,7 +514,7 @@ func (conn *Conn) onICEStateDisconnected(sessionChanged bool) {
 		// todo consider to move after the ConfigureWGEndpoint
 		conn.wgProxyRelay.Work()
 
-		presharedKey := conn.presharedKey(conn.rosenpassRemoteKey, pqPSK)
+		presharedKey := conn.presharedKey(conn.rosenpassRemoteKey)
 		if err := conn.endpointUpdater.SwitchWGEndpoint(conn.wgProxyRelay.EndpointAddr(), presharedKey); err != nil {
 			conn.Log.Errorf("failed to switch to relay conn: %v", err)
 		}
@@ -614,9 +552,6 @@ func (conn *Conn) onICEStateDisconnected(sessionChanged bool) {
 }
 
 func (conn *Conn) onRelayConnectionIsReady(rci RelayConnInfo) {
-	// Read the PQ PSK before conn.mu to keep the lock order conn.mu -> manager.
-	pqPSK, pqOK := conn.pqPSK()
-
 	conn.mu.Lock()
 	defer conn.mu.Unlock()
 
@@ -645,7 +580,7 @@ func (conn *Conn) onRelayConnectionIsReady(rci RelayConnInfo) {
 		conn.Log.Debugf("do not switch to relay because current priority is: %s", conn.currentConnPriority.String())
 		conn.setRelayedProxy(wgProxy)
 		conn.statusRelay.SetConnected()
-		conn.updateRelayStatus(rci.relayedConn.RemoteAddr().String(), rci.rosenpassPubKey, pqOK, time.Now())
+		conn.updateRelayStatus(rci.relayedConn.RemoteAddr().String(), rci.rosenpassPubKey, time.Now())
 		return
 	}
 
@@ -656,7 +591,7 @@ func (conn *Conn) onRelayConnectionIsReady(rci RelayConnInfo) {
 	}
 	updateTime := time.Now()
 	conn.enableWgWatcherIfNeeded(updateTime)
-	if err := conn.endpointUpdater.ConfigureWGEndpoint(wgProxy.EndpointAddr(), conn.presharedKey(rci.rosenpassPubKey, pqPSK)); err != nil {
+	if err := conn.endpointUpdater.ConfigureWGEndpoint(wgProxy.EndpointAddr(), conn.presharedKey(rci.rosenpassPubKey)); err != nil {
 		if err := wgProxy.CloseConn(); err != nil {
 			conn.Log.Warnf("Failed to close relay connection: %v", err)
 		}
@@ -675,7 +610,7 @@ func (conn *Conn) onRelayConnectionIsReady(rci RelayConnInfo) {
 	conn.currentConnPriority = conntype.Relay
 	conn.statusRelay.SetConnected()
 	conn.setRelayedProxy(wgProxy)
-	conn.updateRelayStatus(rci.relayedConn.RemoteAddr().String(), rci.rosenpassPubKey, pqOK, updateTime)
+	conn.updateRelayStatus(rci.relayedConn.RemoteAddr().String(), rci.rosenpassPubKey, updateTime)
 	conn.Log.Infof("start to communicate with peer via relay")
 	conn.doOnConnected(rci.rosenpassPubKey, rci.rosenpassAddr, updateTime)
 }
@@ -737,28 +672,12 @@ func (conn *Conn) onGuardEvent() {
 	}
 }
 
-// RequestReoffer sends a fresh signalling offer for the peer, re-running the
-// post-quantum bootstrap over Signal. Used to recover from a persistent data-path
-// rekey failure: a new exchange overwrites the stalled PSK on both sides. No-op if the
-// connection is not open yet.
-func (conn *Conn) RequestReoffer() {
-	conn.mu.Lock()
-	h := conn.handshaker
-	conn.mu.Unlock()
-	if h == nil {
-		return
-	}
-	if err := h.SendOffer(); err != nil {
-		conn.Log.Debugf("pqkem: recovery re-offer failed: %v", err)
-	}
-}
-
 func (conn *Conn) onWGDisconnected(watcherCtx context.Context) {
 	conn.mu.Lock()
+	defer conn.mu.Unlock()
 
 	// watcherCtx guards against a stale watcher tearing down a connection that already superseded it.
 	if conn.ctx.Err() != nil || watcherCtx.Err() != nil {
-		conn.mu.Unlock()
 		return
 	}
 
@@ -776,15 +695,6 @@ func (conn *Conn) onWGDisconnected(watcherCtx context.Context) {
 	}
 
 	conn.escalateWGTimeoutLocked()
-	pq := conn.config.PQ
-	key := conn.config.Key
-	conn.mu.Unlock()
-
-	// Signal the PQ manager outside conn.mu: it may re-enter Conn (reoffer) under
-	// conn.mu, so calling it while holding the lock would invert the lock order.
-	if pq != nil {
-		pq.OnDataPathDown(key)
-	}
 }
 
 // escalateWGTimeoutLocked resets the peer's rosenpass state after repeated
@@ -808,14 +718,14 @@ func (conn *Conn) escalateWGTimeoutLocked() {
 	conn.onDisconnected(conn.config.WgConfig.RemoteKey)
 }
 
-func (conn *Conn) updateRelayStatus(relayServerAddr string, rosenpassPubKey []byte, pqEstablished bool, updateTime time.Time) {
+func (conn *Conn) updateRelayStatus(relayServerAddr string, rosenpassPubKey []byte, updateTime time.Time) {
 	peerState := State{
 		PubKey:             conn.config.Key,
 		ConnStatusUpdate:   updateTime,
 		ConnStatus:         conn.evalStatus(),
 		Relayed:            conn.isRelayed(),
 		RelayServerAddress: relayServerAddr,
-		RosenpassEnabled:   conn.quantumResistant(rosenpassPubKey, pqEstablished),
+		RosenpassEnabled:   isRosenpassEnabled(rosenpassPubKey),
 	}
 
 	err := conn.statusRecorder.UpdatePeerRelayedState(peerState)
@@ -824,7 +734,7 @@ func (conn *Conn) updateRelayStatus(relayServerAddr string, rosenpassPubKey []by
 	}
 }
 
-func (conn *Conn) updateIceState(iceConnInfo ICEConnInfo, pqEstablished bool, updateTime time.Time) {
+func (conn *Conn) updateIceState(iceConnInfo ICEConnInfo, updateTime time.Time) {
 	peerState := State{
 		PubKey:                     conn.config.Key,
 		ConnStatusUpdate:           updateTime,
@@ -834,7 +744,7 @@ func (conn *Conn) updateIceState(iceConnInfo ICEConnInfo, pqEstablished bool, up
 		RemoteIceCandidateType:     iceConnInfo.RemoteIceCandidateType,
 		LocalIceCandidateEndpoint:  iceConnInfo.LocalIceCandidateEndpoint,
 		RemoteIceCandidateEndpoint: iceConnInfo.RemoteIceCandidateEndpoint,
-		RosenpassEnabled:           conn.quantumResistant(iceConnInfo.RosenpassPubKey, pqEstablished),
+		RosenpassEnabled:           isRosenpassEnabled(iceConnInfo.RosenpassPubKey),
 	}
 
 	err := conn.statusRecorder.UpdatePeerICEState(peerState)
@@ -1038,35 +948,6 @@ func (conn *Conn) onWGCheckSuccess() {
 	conn.mu.Lock()
 	conn.wgTimeouts = 0
 	conn.mu.Unlock()
-
-	// A fresh WireGuard handshake clocks the post-quantum PSK rotation. Pass how long
-	// ago the peer last exchanged real user data (keepalives excluded) so the pqkem
-	// manager can skip rotation on idle tunnels — rotating then would push data-path
-	// traffic that keeps the lazy connection artificially active.
-	if conn.config.PQ != nil {
-		conn.config.PQ.OnDataPathRekeyed(conn.config.Key, conn.dataActivityAge())
-	}
-}
-
-// dataActivityAge returns how long ago the peer last exchanged real user data
-// (WireGuard keepalives excluded), per the same LastActivities signal the
-// lazy-connection inactivity monitor uses. It reports a very large duration when no
-// activity has ever been recorded, so the peer is treated as idle.
-//
-// In kernel mode there is no per-peer data-activity signal (LastActivities is
-// userspace-only), so we cannot tell active from idle. We report zero — always
-// "active" — so PSK rotation is not disabled in kernel mode. Lazy back-to-idle is
-// already limited there; the eBPF WG-activity detection (future) will supply a real
-// signal that excludes handshake/pqkem traffic.
-func (conn *Conn) dataActivityAge() time.Duration {
-	if !conn.config.WgConfig.WgInterface.IsUserspaceBind() {
-		return 0
-	}
-	last, ok := conn.config.WgConfig.WgInterface.LastActivities()[conn.config.WgConfig.RemoteKey]
-	if !ok {
-		return time.Duration(math.MaxInt64)
-	}
-	return monotime.Since(last)
 }
 
 // recordConnectionMetrics records connection stage timestamps as metrics
@@ -1080,12 +961,9 @@ func (conn *Conn) recordConnectionMetrics() {
 	priority := conn.currentConnPriority
 	conn.mu.Unlock()
 
-	var connType metrics.ConnectionType
-	switch priority {
-	case conntype.Relay:
-		connType = metrics.ConnectionTypeRelay
-	default:
-		connType = metrics.ConnectionTypeICE
+	connType := metricsConnType(priority)
+	if connType == metrics.ConnectionTypeUnknown {
+		return
 	}
 
 	// Record metrics with timestamps - duration calculation happens in metrics package
@@ -1107,42 +985,7 @@ func (conn *Conn) AgentVersionString() string {
 	return conn.config.AgentVersion
 }
 
-// pqPSK returns the post-quantum PSK derived for this peer, if the ML-KEM exchange has
-// produced one. It reads the manager WITHOUT conn.mu, so callers fetch it before taking
-// conn.mu: the lock order is always conn.mu -> manager, never the reverse (the manager's
-// reoffer callback re-enters Conn under conn.mu).
-func (conn *Conn) pqPSK() (*wgtypes.Key, bool) {
-	if conn.config.PQ == nil {
-		return nil, false
-	}
-	psk, ok := conn.config.PQ.PSK(conn.config.Key)
-	if !ok {
-		return nil, false
-	}
-	return &psk, true
-}
-
-// presharedKey resolves the WireGuard preshared key for the peer. pqPSK is the
-// post-quantum PSK looked up out of band via pqPSK (nil when none is derived yet), passed
-// in so the manager lock is never taken under conn.mu.
-func (conn *Conn) presharedKey(remoteRosenpassKey []byte, pqPSK *wgtypes.Key) *wgtypes.Key {
-	// Post-quantum: once the ML-KEM exchange has derived a PSK for this peer, program
-	// it here so the peer's next WireGuard handshake adopts it. Applied at peer-config
-	// time (bootstrap / reconnect); steady-state rotation is pushed separately.
-	if conn.config.PQ != nil {
-		if pqPSK != nil {
-			return pqPSK
-		}
-		if conn.config.PQStrict && conn.pqStrictSentinelKey != nil {
-			// Fail closed: program a non-matching sentinel so no session forms on a
-			// non-PQ key until the ML-KEM exchange derives the real PSK (pushed via
-			// SetPresharedKey once it converges). "pending" — turns into a "stuck"
-			// warning from the manager if the exchange keeps failing (see raiseFailure).
-			conn.Log.Debugf("pqkem: strict mode — no PQ PSK yet, blocking peer traffic until the ML-KEM exchange converges")
-			return conn.pqStrictSentinelKey
-		}
-	}
-
+func (conn *Conn) presharedKey(remoteRosenpassKey []byte) *wgtypes.Key {
 	if conn.config.RosenpassConfig.PubKey == nil {
 		return conn.config.WgConfig.PreSharedKey
 	}
@@ -1182,13 +1025,6 @@ func isRosenpassEnabled(remoteRosenpassPubKey []byte) bool {
 	return remoteRosenpassPubKey != nil
 }
 
-// quantumResistant reports whether the peer's tunnel is post-quantum protected, for
-// the status "Quantum resistance" field: either Rosenpass (the remote advertised a
-// Rosenpass key) or the ML-KEM exchange (a PQ PSK has been derived for this peer).
-func (conn *Conn) quantumResistant(remoteRosenpassPubKey []byte, pqEstablished bool) bool {
-	return isRosenpassEnabled(remoteRosenpassPubKey) || pqEstablished
-}
-
 func evalConnStatus(in connStatusInputs) guard.ConnStatus {
 	// "Relay up and needed" — the peer uses relay and the transport is connected.
 	relayUsedAndUp := in.peerUsesRelay && in.relayConnected
@@ -1228,3 +1064,16 @@ func boolToConnStatus(connected bool) guard.ConnStatus {
 	}
 	return guard.ConnStatusDisconnected
 }
+
+func metricsConnType(priority conntype.ConnPriority) metrics.ConnectionType {
+	switch priority {
+	case conntype.Relay:
+		return metrics.ConnectionTypeRelay
+	case conntype.ICETurn:
+		return metrics.ConnectionTypeICETurn
+	case conntype.ICEP2P:
+		return metrics.ConnectionTypeICEP2P
+	default:
+		return metrics.ConnectionTypeUnknown
+	}
+}

client/internal/peer/conn_pq_test.go

@@ -1,90 +0,0 @@
-package peer
-
-import (
-	"net/netip"
-	"testing"
-	"time"
-
-	log "github.com/sirupsen/logrus"
-	"github.com/stretchr/testify/require"
-	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
-)
-
-// fakePQ is a minimal PQHandshaker: only PSK is exercised by presharedKey, the rest
-// are no-op stubs to satisfy the interface.
-type fakePQ struct {
-	psk wgtypes.Key
-	ok  bool
-}
-
-func (f fakePQ) OfferPayload(string) ([]byte, uint16)          { return nil, 0 }
-func (f fakePQ) ShouldSendBootstrapOffer(string) bool          { return false }
-func (f fakePQ) AnswerPayload(string, []byte) ([]byte, uint16) { return nil, 0 }
-func (f fakePQ) OnAnswer(string, []byte)                       {}
-func (f fakePQ) PSK(string) (wgtypes.Key, bool)                { return f.psk, f.ok }
-func (f fakePQ) SetRemoteAddr(string, netip.AddrPort)          {}
-func (f fakePQ) OnDataPathRekeyed(string, time.Duration)       {}
-func (f fakePQ) OnDataPathDown(string)                         {}
-
-// TestConn_presharedKey_PQ covers the post-quantum branch of presharedKey across the
-// three states that matter: a derived PSK is programmed, and — before one exists —
-// strict mode blocks with a sentinel while non-strict falls open to the ordinary key.
-func TestConn_presharedKey_PQ(t *testing.T) {
-	derivedPSK, err := wgtypes.GenerateKey()
-	require.NoError(t, err)
-	nbPSK, err := wgtypes.GenerateKey()
-	require.NoError(t, err)
-
-	newConn := func() *Conn {
-		return &Conn{
-			Log: log.WithField("peer", "pq-test"),
-			config: ConnConfig{
-				Key:             "LLHf3Ma6z6mdLbriAJbqhX7+nM/B71lgw2+91q3LfhU=",
-				LocalKey:        "RRHf3Ma6z6mdLbriAJbqhX7+nM/B71lgw2+91q3LfhU=",
-				WgConfig:        WgConfig{PreSharedKey: &nbPSK},
-				RosenpassConfig: RosenpassConfig{},
-			},
-		}
-	}
-
-	t.Run("derived PSK is programmed", func(t *testing.T) {
-		for _, strict := range []bool{false, true} {
-			c := newConn()
-			c.config.PQ = fakePQ{psk: derivedPSK, ok: true}
-			c.config.PQStrict = strict
-			if strict {
-				sentinel, err := wgtypes.GenerateKey()
-				require.NoError(t, err)
-				c.pqStrictSentinelKey = &sentinel
-			}
-			pqPSK, _ := c.pqPSK()
-			got := c.presharedKey(nil, pqPSK)
-			require.NotNil(t, got)
-			require.Equal(t, derivedPSK, *got, "the derived PQ PSK must win (strict=%v)", strict)
-		}
-	})
-
-	t.Run("non-strict falls open to the ordinary key before a PSK exists", func(t *testing.T) {
-		c := newConn()
-		c.config.PQ = fakePQ{ok: false}
-		c.config.PQStrict = false
-		pqPSK, _ := c.pqPSK()
-		got := c.presharedKey(nil, pqPSK)
-		require.NotNil(t, got, "non-strict must not block")
-		require.Equal(t, nbPSK, *got, "non-strict falls through to the NetBird PSK, not a sentinel")
-	})
-
-	t.Run("strict blocks with the per-conn sentinel before a PSK exists", func(t *testing.T) {
-		sentinel, err := wgtypes.GenerateKey()
-		require.NoError(t, err)
-		c := newConn()
-		c.config.PQ = fakePQ{ok: false}
-		c.config.PQStrict = true
-		c.pqStrictSentinelKey = &sentinel
-		pqPSK, _ := c.pqPSK()
-		got := c.presharedKey(nil, pqPSK)
-		require.NotNil(t, got)
-		require.Equal(t, sentinel, *got, "strict must return the blocking sentinel")
-		require.NotEqual(t, nbPSK, *got, "the sentinel must not be the ordinary key")
-	})
-}

client/internal/peer/conn_test.go

@@ -11,6 +11,8 @@ import (
 	"github.com/stretchr/testify/assert"
 
 	"github.com/netbirdio/netbird/client/iface"
+	"github.com/netbirdio/netbird/client/internal/metrics"
+	"github.com/netbirdio/netbird/client/internal/peer/conntype"
 	"github.com/netbirdio/netbird/client/internal/peer/dispatcher"
 	"github.com/netbirdio/netbird/client/internal/peer/guard"
 	"github.com/netbirdio/netbird/client/internal/peer/ice"
@@ -255,8 +257,8 @@ func TestConn_presharedKey(t *testing.T) {
 			}
 			conn2.config.RosenpassConfig.PermissiveMode = test.conn2Permissive
 
-			conn1PresharedKey := conn1.presharedKey(conn2.config.RosenpassConfig.PubKey, nil)
-			conn2PresharedKey := conn2.presharedKey(conn1.config.RosenpassConfig.PubKey, nil)
+			conn1PresharedKey := conn1.presharedKey(conn2.config.RosenpassConfig.PubKey)
+			conn2PresharedKey := conn2.presharedKey(conn1.config.RosenpassConfig.PubKey)
 
 			if test.conn1ExpectedInitialKey {
 				if conn1PresharedKey == nil {
@@ -294,14 +296,14 @@ func TestConn_presharedKey_RosenpassManaged(t *testing.T) {
 	// When Rosenpass has already initialized the PSK for this peer,
 	// presharedKey must return nil to avoid UpdatePeer overwriting it.
 	conn.rosenpassInitializedPresharedKeyValidator = func(peerKey string) bool { return true }
-	if k := conn.presharedKey([]byte("remote"), nil); k != nil {
+	if k := conn.presharedKey([]byte("remote")); k != nil {
 		t.Fatalf("expected nil presharedKey when Rosenpass manages PSK, got %v", k)
 	}
 
 	// When Rosenpass hasn't taken over yet, presharedKey should provide
 	// a non-nil initial key (deterministic or from NetBird PSK).
 	conn.rosenpassInitializedPresharedKeyValidator = func(peerKey string) bool { return false }
-	if k := conn.presharedKey([]byte("remote"), nil); k == nil {
+	if k := conn.presharedKey([]byte("remote")); k == nil {
 		t.Fatalf("expected non-nil presharedKey before Rosenpass manages PSK")
 	}
 }
@@ -386,3 +388,33 @@ func TestConn_onWGDisconnected_NoEscalationWithoutRosenpass(t *testing.T) {
 	}
 	assert.Empty(t, disconnected, "escalation must be limited to rosenpass connections")
 }
+
+func TestMetricsConnType(t *testing.T) {
+	tests := []struct {
+		name     string
+		priority conntype.ConnPriority
+		expected metrics.ConnectionType
+	}{
+		{"relay", conntype.Relay, metrics.ConnectionTypeRelay},
+		{"ice over turn is relayed, not p2p", conntype.ICETurn, metrics.ConnectionTypeICETurn},
+		{"direct p2p", conntype.ICEP2P, metrics.ConnectionTypeICEP2P},
+		{"unset priority is unknown, not p2p", conntype.None, metrics.ConnectionTypeUnknown},
+		{"unrecognised priority is unknown", conntype.ConnPriority(99), metrics.ConnectionTypeUnknown},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			assert.Equal(t, tc.expected, metricsConnType(tc.priority))
+		})
+	}
+}
+
+func TestMetricsConnType_RelayedMatchesIsRelayed(t *testing.T) {
+	for _, priority := range []conntype.ConnPriority{conntype.None, conntype.Relay, conntype.ICETurn, conntype.ICEP2P} {
+		conn := &Conn{currentConnPriority: priority}
+		tag := metricsConnType(priority)
+		relayedTag := tag == metrics.ConnectionTypeRelay || tag == metrics.ConnectionTypeICETurn
+		assert.Equal(t, conn.isRelayed(), relayedTag,
+			"priority %s: isRelayed and the %q metric tag must agree", priority, tag)
+	}
+}

client/internal/peer/endpoint.go

@@ -88,7 +88,7 @@ func (e *EndpointUpdater) configureAsResponder(addr *net.UDPAddr, presharedKey *
 	var ctx context.Context
 	ctx, e.cancelFunc = context.WithCancel(context.Background())
 	e.updateWg.Add(1)
-	go e.scheduleDelayedUpdate(ctx, addr)
+	go e.scheduleDelayedUpdate(ctx, addr, presharedKey)
 
 	if err := e.updateWireGuardPeer(nil, presharedKey); err != nil {
 		e.waitForCloseTheDelayedUpdate()
@@ -107,14 +107,8 @@ func (e *EndpointUpdater) waitForCloseTheDelayedUpdate() {
 	e.updateWg.Wait()
 }
 
-// scheduleDelayedUpdate waits for the fallback period, then sets the responder's real
-// endpoint. It deliberately passes a nil preshared key so it only updates the endpoint
-// and leaves the current PSK untouched: the PSK captured when this was scheduled may be
-// stale by now (e.g. the post-quantum bootstrap derived a fresher PSK within the
-// fallback window, applied via SetPresharedKey), and re-applying the captured one would
-// revert WireGuard to a key the remote peer no longer uses — a mismatch that stalls the
-// handshake until the next retry.
-func (e *EndpointUpdater) scheduleDelayedUpdate(ctx context.Context, addr *net.UDPAddr) {
+// scheduleDelayedUpdate waits for the fallback period before updating the endpoint
+func (e *EndpointUpdater) scheduleDelayedUpdate(ctx context.Context, addr *net.UDPAddr, presharedKey *wgtypes.Key) {
 	defer e.updateWg.Done()
 	t := time.NewTimer(fallbackDelay)
 	defer t.Stop()
@@ -123,7 +117,7 @@ func (e *EndpointUpdater) scheduleDelayedUpdate(ctx context.Context, addr *net.U
 	case <-ctx.Done():
 		return
 	case <-t.C:
-		if err := e.updateWireGuardPeer(addr, nil); err != nil {
+		if err := e.updateWireGuardPeer(addr, presharedKey); err != nil {
 			e.log.Errorf("failed to update WireGuard peer, address: %s, error: %v", addr, err)
 		}
 	}

client/internal/peer/handshaker.go

@@ -39,16 +39,6 @@ type OfferAnswer struct {
 	// This value is the local Rosenpass server address when sending the message
 	RosenpassAddr string
 
-	// MlkemPayload carries the post-quantum X25519MLKEM768 handshake message
-	// (pqkem-framed offer on an OFFER, answer on an ANSWER) that seeds the
-	// WireGuard PSK. Opaque here — the pqkem library frames and parses it. Nil
-	// when the peer does not run the ML-KEM PQ exchange.
-	MlkemPayload []byte
-
-	// MlkemPort is the peer's ML-KEM PQ service UDP port (bound on its WG overlay
-	// IP) where data-path rekey messages are sent. Zero when not running the exchange.
-	MlkemPort uint16
-
 	// relay server address
 	RelaySrvAddress string
 	// RelaySrvIP is the IP the remote peer is connected to on its
@@ -91,20 +81,14 @@ type Handshaker struct {
 
 func NewHandshaker(log *log.Entry, config ConnConfig, signaler *Signaler, ice *WorkerICE, relay *WorkerRelay, metricsStages *MetricsStages) *Handshaker {
 	h := &Handshaker{
-		log:           log,
-		config:        config,
-		signaler:      signaler,
-		ice:           ice,
-		relay:         relay,
-		metricsStages: metricsStages,
-		// Buffered by 1: the single Listen goroutine can be busy handling an offer
-		// (sendAnswer does a blocking signal send) exactly when the matching answer
-		// arrives on the other channel. Unbuffered, that answer would hit the
-		// non-blocking send's default and be dropped — fatal for the post-quantum
-		// exchange, which needs the answer to converge. A 1-slot cushion lets it wait
-		// until Listen loops back, without ever blocking the signal receiver.
-		remoteOffersCh: make(chan OfferAnswer, 1),
-		remoteAnswerCh: make(chan OfferAnswer, 1),
+		log:            log,
+		config:         config,
+		signaler:       signaler,
+		ice:            ice,
+		relay:          relay,
+		metricsStages:  metricsStages,
+		remoteOffersCh: make(chan OfferAnswer),
+		remoteAnswerCh: make(chan OfferAnswer),
 	}
 	// assume remote supports ICE until we learn otherwise from received offers
 	h.remoteICESupported.Store(ice != nil)
@@ -127,9 +111,44 @@ func (h *Handshaker) Listen(ctx context.Context) {
 	for {
 		select {
 		case remoteOfferAnswer := <-h.remoteOffersCh:
-			h.handleRemoteOffer(remoteOfferAnswer)
+			h.log.Infof("received offer, running version %s, remote WireGuard listen port %d, session id: %s, remote ICE supported: %t", remoteOfferAnswer.Version, remoteOfferAnswer.WgListenPort, remoteOfferAnswer.SessionIDString(), remoteOfferAnswer.hasICECredentials())
+
+			// Record signaling received for reconnection attempts
+			if h.metricsStages != nil {
+				h.metricsStages.RecordSignalingReceived()
+			}
+
+			h.updateRemoteICEState(&remoteOfferAnswer)
+
+			if h.relayListener != nil {
+				h.relayListener.Notify(&remoteOfferAnswer)
+			}
+
+			if h.iceListener != nil && h.RemoteICESupported() {
+				h.iceListener(&remoteOfferAnswer)
+			}
+
+			if err := h.sendAnswer(); err != nil {
+				h.log.Errorf("failed to send remote offer confirmation: %s", err)
+				continue
+			}
 		case remoteOfferAnswer := <-h.remoteAnswerCh:
-			h.handleRemoteAnswer(remoteOfferAnswer)
+			h.log.Infof("received answer, running version %s, remote WireGuard listen port %d, session id: %s, remote ICE supported: %t", remoteOfferAnswer.Version, remoteOfferAnswer.WgListenPort, remoteOfferAnswer.SessionIDString(), remoteOfferAnswer.hasICECredentials())
+
+			// Record signaling received for reconnection attempts
+			if h.metricsStages != nil {
+				h.metricsStages.RecordSignalingReceived()
+			}
+
+			h.updateRemoteICEState(&remoteOfferAnswer)
+
+			if h.relayListener != nil {
+				h.relayListener.Notify(&remoteOfferAnswer)
+			}
+
+			if h.iceListener != nil && h.RemoteICESupported() {
+				h.iceListener(&remoteOfferAnswer)
+			}
 		case <-ctx.Done():
 			h.log.Infof("stop listening for remote offers and answers")
 			return
@@ -137,119 +156,6 @@ func (h *Handshaker) Listen(ctx context.Context) {
 	}
 }
 
-// onSignalReceived runs the common preamble for a received offer/answer: record the
-// signalling metric, refresh the remote ICE state, and register the peer's post-quantum
-// data-path endpoint learned from the message.
-func (h *Handshaker) onSignalReceived(remoteOfferAnswer *OfferAnswer) {
-	if h.metricsStages != nil {
-		h.metricsStages.RecordSignalingReceived()
-	}
-	h.updateRemoteICEState(remoteOfferAnswer)
-	h.pqRegisterEndpoint(remoteOfferAnswer.MlkemPort)
-}
-
-// notifyListeners hands the offer/answer to the relay and ICE workers so they bring the
-// connection up.
-func (h *Handshaker) notifyListeners(remoteOfferAnswer *OfferAnswer) {
-	if h.relayListener != nil {
-		h.relayListener.Notify(remoteOfferAnswer)
-	}
-	if h.iceListener != nil && h.RemoteICESupported() {
-		h.iceListener(remoteOfferAnswer)
-	}
-}
-
-func (h *Handshaker) handleRemoteOffer(remoteOfferAnswer OfferAnswer) {
-	h.log.Infof("received offer, running version %s, remote WireGuard listen port %d, session id: %s, remote ICE supported: %t", remoteOfferAnswer.Version, remoteOfferAnswer.WgListenPort, remoteOfferAnswer.SessionIDString(), remoteOfferAnswer.hasICECredentials())
-	h.onSignalReceived(&remoteOfferAnswer)
-
-	// If we are the controller running the KEM, a responder's offer is handled by
-	// replying with our own KEM offer, not by answering it (see pqControllerReoffer).
-	if h.pqControllerReoffer() {
-		return
-	}
-
-	// Derive+store the KEM PSK (inside sendAnswer's AnswerPayload) BEFORE bringing up the
-	// connection: the relay/ICE workers configure the WG endpoint, which pulls the PSK
-	// for the first handshake. Notifying them first would race the KEM exchange and hand
-	// the first handshake a not-yet-derived key.
-	if err := h.sendAnswer(&remoteOfferAnswer); err != nil {
-		h.log.Errorf("failed to send remote offer confirmation: %s", err)
-		return
-	}
-	h.notifyListeners(&remoteOfferAnswer)
-}
-
-func (h *Handshaker) handleRemoteAnswer(remoteOfferAnswer OfferAnswer) {
-	h.log.Infof("received answer, running version %s, remote WireGuard listen port %d, session id: %s, remote ICE supported: %t", remoteOfferAnswer.Version, remoteOfferAnswer.WgListenPort, remoteOfferAnswer.SessionIDString(), remoteOfferAnswer.hasICECredentials())
-	h.onSignalReceived(&remoteOfferAnswer)
-
-	// Feed the KEM answer (derive+store PSK) BEFORE bringing up the connection so the WG
-	// endpoint config pulls the real PSK for the first handshake instead of racing ahead
-	// of the KEM exchange.
-	if h.config.PQ != nil {
-		h.config.PQ.OnAnswer(h.config.Key, remoteOfferAnswer.MlkemPayload)
-	}
-	h.notifyListeners(&remoteOfferAnswer)
-}
-
-// pqControllerReoffer handles a responder's offer when we are the controller running the
-// KEM. The KEM material rides only the controller's offer, so the two peers derive a
-// single shared PSK (a bidirectional KEM would yield two different PSKs and WireGuard
-// would pick misaligned ones). Rather than answer the responder's (KEM-less) offer —
-// which would bring WireGuard up on a pre-PQ key before the KEM completes — we reply with
-// our own KEM offer, so the only transaction that establishes the tunnel is the one that
-// also derives the PSK. It also guarantees a responder-initiated wake still triggers a
-// KEM offer (no stuck responder). Sent exactly once per exchange; further offers while
-// one is in flight are ignored (re-sending on every responder offer would be a runaway).
-// The re-offer reuses our stable ICE session id, so the peer dedups repeats.
-//
-// Returns true when it took ownership of the offer (the caller must not answer it).
-func (h *Handshaker) pqControllerReoffer() bool {
-	if h.config.PQ == nil || !isController(h.config) {
-		return false
-	}
-	if h.config.PQ.ShouldSendBootstrapOffer(h.config.Key) {
-		h.log.Debugf("pqkem: controller received a responder offer, replying with our KEM offer instead of an answer")
-		if err := h.sendOffer(); err != nil {
-			h.log.Errorf("failed to send KEM offer in response to peer offer: %s", err)
-		}
-	} else {
-		h.log.Debugf("pqkem: controller received a responder offer but a KEM exchange is already in flight, ignoring")
-	}
-	return true
-}
-
-// pqRegisterEndpoint feeds the post-quantum handshaker the peer's data-path endpoint
-// (its WG overlay IP plus the advertised pq UDP port) learned from a remote offer/answer.
-func (h *Handshaker) pqRegisterEndpoint(remotePort uint16) {
-	if h.config.PQ == nil {
-		return
-	}
-	overlay, ok := h.pqPeerOverlayAddr()
-	if !ok {
-		return
-	}
-	// remotePort may be 0 (the peer omitted it, meaning the default port); the adapter
-	// resolves 0 to DefaultPort.
-	h.config.PQ.SetRemoteAddr(h.config.Key, netip.AddrPortFrom(overlay, remotePort))
-}
-
-// pqPeerOverlayAddr returns the peer's IPv4 overlay address for the pq data path. A
-// RemotePeerConfig carries only the peer overlay (v4 /32, optionally v6 /128) — served
-// routes are programmed on WireGuard separately and never land in WgConfig.AllowedIps —
-// so AllowedIps[0] is the v4 overlay, matching conn.AllowedIP(). The transport is v4
-// (the overlay always has v4; v6 is additive). Returns false when no v4 overlay exists.
-func (h *Handshaker) pqPeerOverlayAddr() (netip.Addr, bool) {
-	if len(h.config.WgConfig.AllowedIps) == 0 {
-		return netip.Addr{}, false
-	}
-	if a := h.config.WgConfig.AllowedIps[0].Addr().Unmap(); a.Is4() {
-		return a, true
-	}
-	return netip.Addr{}, false
-}
-
 func (h *Handshaker) SendOffer() error {
 	h.mu.Lock()
 	defer h.mu.Unlock()
@@ -289,23 +195,13 @@ func (h *Handshaker) sendOffer() error {
 	}
 
 	offer := h.buildOfferAnswer()
-	if h.config.PQ != nil {
-		offer.MlkemPayload, offer.MlkemPort = h.config.PQ.OfferPayload(h.config.Key)
-	}
 	h.log.Debugf("sending offer with serial: %s", offer.SessionIDString())
 
 	return h.signaler.SignalOffer(offer, h.config.Key)
 }
 
-func (h *Handshaker) sendAnswer(remoteOffer *OfferAnswer) error {
+func (h *Handshaker) sendAnswer() error {
 	answer := h.buildOfferAnswer()
-	if h.config.PQ != nil {
-		var recvOffer []byte
-		if remoteOffer != nil {
-			recvOffer = remoteOffer.MlkemPayload
-		}
-		answer.MlkemPayload, answer.MlkemPort = h.config.PQ.AnswerPayload(h.config.Key, recvOffer)
-	}
 	h.log.Debugf("sending answer with serial: %s", answer.SessionIDString())
 
 	return h.signaler.SignalAnswer(answer, h.config.Key)

client/internal/peer/iface.go

@@ -10,7 +10,6 @@ import (
 	"github.com/netbirdio/netbird/client/iface/configurer"
 	"github.com/netbirdio/netbird/client/iface/wgaddr"
 	"github.com/netbirdio/netbird/client/iface/wgproxy"
-	"github.com/netbirdio/netbird/monotime"
 )
 
 type WGIface interface {
@@ -20,11 +19,4 @@ type WGIface interface {
 	GetProxy() wgproxy.Proxy
 	Address() wgaddr.Address
 	RemoveEndpointAddress(key string) error
-	// LastActivities returns the last real-data activity time per peer (WireGuard
-	// keepalives excluded), used to gate post-quantum PSK rotation on active tunnels.
-	LastActivities() map[string]monotime.Time
-	// IsUserspaceBind reports whether WireGuard runs in userspace. Only there does
-	// LastActivities track per-peer data activity; in kernel mode it is unavailable,
-	// so PSK rotation cannot be gated on activity.
-	IsUserspaceBind() bool
 }

client/internal/peer/signaler.go

@@ -63,8 +63,6 @@ func (s *Signaler) signalOfferAnswer(offerAnswer OfferAnswer, remoteKey string,
 		},
 		RosenpassPubKey: offerAnswer.RosenpassPubKey,
 		RosenpassAddr:   offerAnswer.RosenpassAddr,
-		MlkemPayload:    offerAnswer.MlkemPayload,
-		MlkemPort:       int(offerAnswer.MlkemPort),
 		RelaySrvAddress: offerAnswer.RelaySrvAddress,
 		RelaySrvIP:      offerAnswer.RelaySrvIP,
 		SessionID:       sessionIDBytes,

client/internal/pqkem/bench_test.go

@@ -1,71 +0,0 @@
-package pqkem
-
-import (
-	"crypto/ecdh"
-	"crypto/mlkem"
-	"crypto/rand"
-	"testing"
-)
-
-func BenchmarkX25519Keygen(b *testing.B) {
-	c := ecdh.X25519()
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		if _, err := c.GenerateKey(rand.Reader); err != nil {
-			b.Fatal(err)
-		}
-	}
-}
-
-func BenchmarkX25519ECDH(b *testing.B) {
-	c := ecdh.X25519()
-	a, err := c.GenerateKey(rand.Reader)
-	if err != nil {
-		b.Fatal(err)
-	}
-	p, err := c.GenerateKey(rand.Reader)
-	if err != nil {
-		b.Fatal(err)
-	}
-	pub := p.PublicKey()
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		if _, err := a.ECDH(pub); err != nil {
-			b.Fatal(err)
-		}
-	}
-}
-
-func BenchmarkMLKEMKeygen(b *testing.B) {
-	for i := 0; i < b.N; i++ {
-		if _, err := mlkem.GenerateKey768(); err != nil {
-			b.Fatal(err)
-		}
-	}
-}
-
-func BenchmarkMLKEMEncaps(b *testing.B) {
-	dk, err := mlkem.GenerateKey768()
-	if err != nil {
-		b.Fatal(err)
-	}
-	ek := dk.EncapsulationKey()
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		_, _ = ek.Encapsulate()
-	}
-}
-
-func BenchmarkMLKEMDecaps(b *testing.B) {
-	dk, err := mlkem.GenerateKey768()
-	if err != nil {
-		b.Fatal(err)
-	}
-	_, ct := dk.EncapsulationKey().Encapsulate()
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		if _, err := dk.Decapsulate(ct); err != nil {
-			b.Fatal(err)
-		}
-	}
-}

client/internal/pqkem/callbacks.go

@@ -1,18 +0,0 @@
-package pqkem
-
-// CallbackHandler is implemented by the host and invoked by the library. The
-// library only reports events; the host owns the reaction. Keeping this an
-// interface — rather than touching the transport or keying directly — is what lets
-// the KEM code be extracted as a standalone library.
-type CallbackHandler interface {
-	// OnNewPSKReady fires when a fresh post-quantum PSK has been derived for a peer
-	// and must be programmed into the consumer's secure channel. It is invoked at
-	// the commit point of each side: the initiator on receiving the answer, the
-	// responder on receiving the confirm.
-	OnNewPSKReady(remoteID RemoteID, psk PSK) error
-
-	// OnRekeyFailed fires when an exchange fails to converge within the allotted
-	// time. The host should tear the peer connection down so it re-establishes, and
-	// log a WARN. The library reports the event; it does not dictate the reaction.
-	OnRekeyFailed(remoteID RemoteID) error
-}

client/internal/pqkem/capability_test.go

@@ -1,66 +0,0 @@
-package pqkem
-
-import (
-	"testing"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// TestManager_NonCapablePeerNotOffered: a peer known not to run the KEM (it advertised
-// no PQ port over signalling) is never offered an exchange, and no failure is raised —
-// this is what stops the reoffer storm against non-PQ peers (e.g. Rosenpass peers).
-func TestManager_NonCapablePeerNotOffered(t *testing.T) {
-	wg := newFakeWG()
-	d := NewManager("bbbb", wg, nil) // initiator vs "aaaa"
-	d.Start(&loopback{ep: epB, sw: newSwitch()})
-	defer d.Stop()
-
-	d.MarkNonCapable("aaaa")
-
-	offer, err := d.SignalOffer("aaaa")
-	require.NoError(t, err)
-	assert.Nil(t, offer, "a non-capable peer must not be offered a KEM exchange")
-	assert.Empty(t, wg.failed, "a non-capable peer must not raise a rekey failure")
-}
-
-// TestManager_MarkNonCapableCancelsInFlight: if we start an exchange with a peer whose
-// capability is not yet known and then learn it does not run the KEM, the in-flight
-// exchange is cancelled and no further offer is produced (no timeout -> no failure).
-func TestManager_MarkNonCapableCancelsInFlight(t *testing.T) {
-	wg := newFakeWG()
-	d := NewManager("bbbb", wg, nil)
-	d.Start(&loopback{ep: epB, sw: newSwitch()})
-	defer d.Stop()
-
-	// Capability unknown -> the bootstrap offer goes out optimistically.
-	offer, err := d.SignalOffer("aaaa")
-	require.NoError(t, err)
-	require.NotNil(t, offer)
-
-	// Now we learn the peer is non-PQ: the exchange must be dropped.
-	d.MarkNonCapable("aaaa")
-
-	next, err := d.SignalOffer("aaaa")
-	require.NoError(t, err)
-	assert.Nil(t, next, "after learning non-capability the peer is no longer offered")
-	assert.Empty(t, wg.failed, "cancelling an in-flight exchange must not raise a failure")
-}
-
-// TestManager_EstablishedPeerNotDowngraded: a stray zero-port observation must not tear
-// down a peer we already have a working PQ session with.
-func TestManager_EstablishedPeerNotDowngraded(t *testing.T) {
-	dA, dB, _, wgB, _ := pair(t)
-	defer dA.Stop()
-	defer dB.Stop()
-
-	bootstrap(t, dA, dB)
-	require.NotEqual(t, PSK{}, wgB.psk("aaaa"), "established a PSK")
-
-	dB.MarkNonCapable("aaaa") // stray zero after establishment
-
-	// The peer keeps its derived PSK (MarkNonCapable is a no-op once established).
-	psk, ok := dB.PSK("aaaa")
-	assert.True(t, ok, "an established peer must keep its PSK despite a stray zero")
-	assert.NotEqual(t, PSK{}, psk)
-}

client/internal/pqkem/concurrency_test.go

@@ -1,77 +0,0 @@
-package pqkem
-
-import (
-	"sync"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/require"
-)
-
-// TestConcurrency_RecoversViaResignalAfterDataPathBreak exercises the A-light recovery:
-// a data-path rotation can no longer converge (OnRekeyFailed), and re-bootstrapping over
-// signalling resyncs both peers on a fresh PSK — even while the data path stays broken,
-// since the signal channel is independent of it.
-func TestConcurrency_RecoversViaResignalAfterDataPathBreak(t *testing.T) {
-	dA, dB, wgA, wgB, lbB := pair(t)
-	defer dA.Stop()
-	defer dB.Stop()
-
-	// Tighten B's timings and make a single rotation miss raise OnRekeyFailed. Set
-	// before any exchange loop spawns (the loop reads these fields).
-	dB.retryInterval = 5 * time.Millisecond
-	dB.maxRetries = 2
-	dB.maxRekeyFailures = 1
-
-	bootstrap(t, dA, dB)
-	dA.OnDataPathRekeyed("bbbb", 0)
-	dB.OnDataPathRekeyed("aaaa", 0)
-	psk1 := wgB.psk("aaaa")
-	require.NotEqual(t, PSK{}, psk1)
-	require.Equal(t, psk1, wgA.psk("bbbb"), "converged on the same PSK after bootstrap+rotation")
-
-	// Data path breaks: the rotation can no longer converge -> OnRekeyFailed.
-	lbB.drop.Store(true)
-	_, err := dB.startExchangeTest("aaaa", false, ExchangeID{})
-	require.NoError(t, err)
-	require.Eventually(t, func() bool { return failedCount(wgB) >= 1 }, time.Second, 5*time.Millisecond)
-
-	// Recovery: re-bootstrap over signalling with the data path STILL broken. It must
-	// still converge (signal is independent of the data path) on a fresh PSK.
-	bootstrap(t, dA, dB)
-	psk2 := wgB.psk("aaaa")
-	require.NotEqual(t, psk1, psk2, "recovery derived a fresh PSK")
-	require.Equal(t, psk2, wgA.psk("bbbb"), "both sides resync after recovery")
-}
-
-// TestConcurrency_ConcurrentRekeysNoRace hammers both managers with concurrent rotation
-// clocks from many goroutines. Its primary job (with -race) is to prove the single-lock
-// state machine has no data races or deadlocks under contention; a final deterministic
-// bootstrap then asserts there is no split-brain (both sides on the same PSK).
-func TestConcurrency_ConcurrentRekeysNoRace(t *testing.T) {
-	dA, dB, wgA, wgB, _ := pair(t)
-	defer dA.Stop()
-	defer dB.Stop()
-
-	bootstrap(t, dA, dB)
-
-	var wg sync.WaitGroup
-	for g := 0; g < 8; g++ {
-		wg.Add(1)
-		go func() {
-			defer wg.Done()
-			for i := 0; i < 50; i++ {
-				dB.OnDataPathRekeyed("aaaa", 0) // initiator chains a rotation
-				dA.OnDataPathRekeyed("bbbb", 0) // responder side is a no-op, still stresses the lock
-			}
-		}()
-	}
-	wg.Wait()
-
-	// The storm may leave an exchange mid-flight (concurrent cancellation). Force a
-	// clean convergence over signalling, then assert no split-brain.
-	bootstrap(t, dA, dB)
-	a, b := wgA.psk("bbbb"), wgB.psk("aaaa")
-	require.NotEqual(t, PSK{}, b)
-	require.Equal(t, a, b, "both sides converge on the same PSK, no split-brain")
-}

client/internal/pqkem/convergence.go

@@ -1,291 +0,0 @@
-package pqkem
-
-import (
-	"context"
-	"crypto/sha256"
-	"encoding/hex"
-	"fmt"
-	"time"
-)
-
-// idHex renders an exchange ID for logs.
-func idHex(id ExchangeID) string { return hex.EncodeToString(id[:]) }
-
-// pskFingerprint is a short, non-secret digest of a derived PSK: identical on both
-// peers iff they derived the same key. Logged instead of the raw PSK so debug logs
-// never carry the actual WireGuard preshared key.
-func pskFingerprint(psk PSK) string {
-	sum := sha256.Sum256(psk[:])
-	return hex.EncodeToString(sum[:8])
-}
-
-// startExchange creates a fresh initiator exchange (acknowledging ackID, zero for a
-// bootstrap) and returns the framed offer for the caller to send — pushed over the
-// data path for a chained rekey, or handed to the host for signalling when viaSignal
-// is set. Any previous in-flight exchange for the peer is cancelled.
-// startExchangeLocked must be called with m.mu held: the caller's idempotency check and
-// the exchange install stay under one lock acquisition so two concurrent starts for the
-// same peer cannot both create an exchange. It also refuses to start (and to Add to the
-// wait group) once the manager is stopping, so it never races Manager.Stop's Wait.
-func (m *Manager) startExchangeLocked(remoteID RemoteID, viaSignal bool, ackID ExchangeID) ([]byte, error) {
-	if m.rootCtx.Err() != nil {
-		return nil, fmt.Errorf("manager stopping")
-	}
-	init, err := NewInitiator()
-	if err != nil {
-		return nil, err
-	}
-	id, err := newExchangeID()
-	if err != nil {
-		return nil, err
-	}
-	raw, err := (&OfferMsg{ExchangeID: id, AckID: ackID, KEMOffer: init.Offer()}).Encode()
-	if err != nil {
-		return nil, err
-	}
-
-	ctx, cancel := context.WithCancel(m.rootCtx)
-	if old := m.exchanges[remoteID]; old != nil && old.cancel != nil {
-		old.cancel()
-	}
-	m.exchanges[remoteID] = &exchangeCtl{
-		id:        id,
-		state:     stateAwaitingAnswer,
-		startedAt: time.Now(),
-		cancel:    cancel,
-		lastSent:  raw,
-		initiator: init,
-		viaSignal: viaSignal,
-	}
-
-	m.wait.Add(1)
-	go m.initiatorLoop(ctx, remoteID, id)
-
-	via := "data-path"
-	if viaSignal {
-		via = "signal"
-	}
-	m.trace("pqkem: offer sent", "peer", remoteID, "exchange", idHex(id), "acks", idHex(ackID), "via", via)
-	return raw, nil
-}
-
-// processOffer (responder) first acknowledges the previous exchange the offer names
-// (that offer riding the data path under the freshly adopted key proves it worked),
-// then derives the PSK for the new offer, commits it optimistically, and returns the
-// framed answer. A duplicate offer returns the cached answer without re-deriving.
-func (m *Manager) processOffer(remoteID RemoteID, o *OfferMsg) ([]byte, error) {
-	m.trace("pqkem: offer received", "peer", remoteID, "exchange", idHex(o.ExchangeID), "acks", idHex(o.AckID))
-
-	if o.AckID != (ExchangeID{}) {
-		m.ackConverged(remoteID, o.AckID)
-	}
-
-	m.mu.Lock()
-	if ex := m.exchanges[remoteID]; ex != nil && ex.id == o.ExchangeID {
-		state, last := ex.state, ex.lastSent
-		m.mu.Unlock()
-		if state == stateReserved {
-			return nil, nil
-		}
-		m.trace("pqkem: duplicate offer, resending cached answer", "peer", remoteID, "exchange", idHex(o.ExchangeID))
-		return last, nil
-	}
-	// Reserve the slot so a concurrent duplicate offer bails.
-	m.exchanges[remoteID] = &exchangeCtl{id: o.ExchangeID, state: stateReserved, startedAt: time.Now()}
-	m.mu.Unlock()
-
-	answerBytes, psk, err := Respond(o.KEMOffer, m.binding(remoteID))
-	if err != nil {
-		return nil, err
-	}
-	raw, err := (&AnswerMsg{ExchangeID: o.ExchangeID, KEMAnswer: answerBytes}).Encode()
-	if err != nil {
-		return nil, err
-	}
-
-	m.mu.Lock()
-	ex := m.exchanges[remoteID]
-	if ex == nil || ex.id != o.ExchangeID {
-		m.mu.Unlock()
-		m.trace("pqkem: exchange superseded during respond, dropping answer", "peer", remoteID, "exchange", idHex(o.ExchangeID))
-		return nil, nil
-	}
-	ex.state = stateAwaitingAck
-	ex.lastSent = raw
-	ex.pendingPSK = psk
-	m.psks[remoteID] = psk
-	m.capable[remoteID] = true // a real KEM offer proves the peer runs the exchange
-	m.mu.Unlock()
-
-	m.trace("pqkem: new PSK derived", "peer", remoteID, "exchange", idHex(o.ExchangeID), "role", "responder", "psk_fp", pskFingerprint(psk))
-
-	// Commit optimistically so our data path can rekey to the new PSK.
-	if err := m.cbHandler.OnNewPSKReady(remoteID, psk); err != nil {
-		return nil, err
-	}
-	m.trace("pqkem: answer sent", "peer", remoteID, "exchange", idHex(o.ExchangeID))
-	return raw, nil
-}
-
-// processAnswer (initiator) derives and commits the PSK and parks in
-// stateAwaitingRekey; the next offer (chained from OnDataPathRekeyed) will acknowledge
-// this exchange. Only valid in stateAwaitingAnswer; advancing the state under the
-// lock makes a concurrent/duplicate answer bail.
-func (m *Manager) processAnswer(remoteID RemoteID, a *AnswerMsg) error {
-	m.mu.Lock()
-	ex := m.exchanges[remoteID]
-	if ex == nil || ex.id != a.ExchangeID || ex.state != stateAwaitingAnswer {
-		haveID := "none"
-		if ex != nil {
-			haveID = idHex(ex.id)
-		}
-		m.mu.Unlock()
-		m.trace("pqkem: unexpected answer dropped (inconsistency)", "peer", remoteID, "answer_for", idHex(a.ExchangeID), "have_exchange", haveID)
-		return nil
-	}
-	ex.state = stateAwaitingRekey
-	init := ex.initiator
-	ex.initiator = nil
-	m.mu.Unlock()
-
-	m.trace("pqkem: answer received", "peer", remoteID, "exchange", idHex(a.ExchangeID))
-
-	psk, err := init.Finish(a.KEMAnswer, m.binding(remoteID))
-	if err != nil {
-		// The state already advanced to stateAwaitingRekey and the initiator was cleared,
-		// so initiatorLoop would exit its default branch without registering a failure —
-		// leaving the peer desynced (the responder committed its PSK in processOffer).
-		// Drop the exchange and raise the failure so recovery re-bootstraps.
-		m.mu.Lock()
-		if cur := m.exchanges[remoteID]; cur != nil && cur.id == a.ExchangeID {
-			delete(m.exchanges, remoteID)
-		}
-		initial := !m.established[remoteID]
-		fail := m.registerFailureLocked(remoteID)
-		m.mu.Unlock()
-		m.raiseFailure(remoteID, fail, initial)
-		return err
-	}
-
-	// The initiator has converged: the responder must have derived the key to answer.
-	m.mu.Lock()
-	m.established[remoteID] = true
-	m.failures[remoteID] = 0
-	m.psks[remoteID] = psk
-	m.capable[remoteID] = true // a real KEM answer proves the peer runs the exchange
-	m.mu.Unlock()
-
-	m.trace("pqkem: new PSK derived", "peer", remoteID, "exchange", idHex(a.ExchangeID), "role", "initiator", "psk_fp", pskFingerprint(psk))
-
-	return m.cbHandler.OnNewPSKReady(remoteID, psk)
-}
-
-// ackConverged (responder) records convergence of the exchange named by ackID: a
-// later offer acknowledging it proves both sides operate on that exchange's key. Only
-// acts on a matching stateAwaitingAck exchange; anything else is ignored.
-func (m *Manager) ackConverged(remoteID RemoteID, ackID ExchangeID) {
-	m.mu.Lock()
-	ex := m.exchanges[remoteID]
-	if ex == nil || ex.id != ackID || ex.state != stateAwaitingAck {
-		m.mu.Unlock()
-		m.trace("pqkem: ack for unknown/mismatched exchange, ignored (inconsistency)", "peer", remoteID, "acks", idHex(ackID))
-		return
-	}
-	delete(m.exchanges, remoteID)
-	m.established[remoteID] = true
-	m.failures[remoteID] = 0
-	_ = time.Since(ex.startedAt) // convergence latency (metrics hook, later step)
-	m.mu.Unlock()
-
-	m.trace("pqkem: previous exchange confirmed by ack", "peer", remoteID, "exchange", idHex(ackID))
-}
-
-// initiatorLoop enforces the offer->answer convergence deadline and retransmits the
-// initiator's outstanding data-path offer while awaiting the answer (a
-// signalling-bootstrapped offer is retransmitted by the host, so it is not resent
-// here). Exhausting the deadline before the answer arrives is a failure. Once the
-// answer is in (state past awaitingAnswer) the loop exits: the next rotation is driven
-// by OnDataPathRekeyed, and the idle wait for it has no deadline.
-func (m *Manager) initiatorLoop(ctx context.Context, remoteID RemoteID, id ExchangeID) {
-	defer m.wait.Done()
-	t := time.NewTicker(m.retryInterval)
-	defer t.Stop()
-
-	attempts := 0
-	for {
-		select {
-		case <-ctx.Done():
-			return
-		case <-t.C:
-			m.mu.Lock()
-			ex := m.exchanges[remoteID]
-			if ex == nil || ex.id != id {
-				m.mu.Unlock()
-				return
-			}
-
-			switch ex.state {
-			case stateAwaitingAnswer:
-				if attempts >= m.maxRetries {
-					delete(m.exchanges, remoteID)
-					initial := !m.established[remoteID]
-					fail := m.registerFailureLocked(remoteID)
-					m.mu.Unlock()
-					m.raiseFailure(remoteID, fail, initial)
-					return
-				}
-				viaSignal := ex.viaSignal
-				msg := ex.lastSent
-				attempts++
-				m.mu.Unlock()
-				if !viaSignal {
-					if err := m.pushDataPath(remoteID, msg); err != nil {
-						m.logger.Warn("pqkem: offer retransmit failed", "peer", remoteID, "err", err)
-					}
-				}
-
-			default:
-				// Past awaiting the answer (converged) or superseded: the loop's job
-				// is done. The next rotation is driven externally by OnDataPathRekeyed,
-				// so there is no deadline while idle-waiting for it (that wait can be
-				// as long as the transport's natural rekey interval).
-				m.mu.Unlock()
-				return
-			}
-		}
-	}
-}
-
-// registerFailureLocked applies policy B and reports whether OnRekeyFailed is due:
-// an initial exchange (peer never established) fails immediately; a rekey tolerates
-// up to maxRekeyFailures consecutive misses (we stay on the still-valid previous
-// PSK) before failing. Assumes m.mu is held.
-func (m *Manager) registerFailureLocked(remoteID RemoteID) bool {
-	if !m.established[remoteID] {
-		return true
-	}
-	m.failures[remoteID]++
-	if m.failures[remoteID] >= m.maxRekeyFailures {
-		m.failures[remoteID] = 0
-		return true
-	}
-	return false
-}
-
-// raiseFailure reports a convergence failure. initial distinguishes a never-established
-// peer (bootstrap failed → no PQ PSK at all; in strict mode the peer stays blocked =
-// "stuck") from a rekey failure (a previous PSK is still in force and traffic continues).
-func (m *Manager) raiseFailure(remoteID RemoteID, fail, initial bool) {
-	if !fail {
-		m.logger.Warn("pqkem: rekey attempt timed out, will retry next cycle", "peer", remoteID)
-		return
-	}
-	if initial {
-		m.logger.Warn("pqkem: initial exchange failed — no PQ PSK established for peer (strict mode keeps the peer blocked until it converges)", "peer", remoteID)
-	} else {
-		m.logger.Warn("pqkem: rekey failed after retries — staying on the previous PSK", "peer", remoteID)
-	}
-	if err := m.cbHandler.OnRekeyFailed(remoteID); err != nil {
-		m.logger.Error("pqkem: OnRekeyFailed handler error", "peer", remoteID, "err", err)
-	}
-}

client/internal/pqkem/convergence_test.go

@@ -1,75 +0,0 @@
-package pqkem
-
-import (
-	"net/netip"
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/assert"
-	"github.com/stretchr/testify/require"
-)
-
-// dropTransport is a pqkem.Transport that silently discards everything.
-type dropTransport struct{}
-
-func (dropTransport) Send(netip.AddrPort, []byte) error { return nil }
-func (dropTransport) LocalPort() int                    { return 0 }
-func (dropTransport) Run(func(netip.AddrPort, []byte))  {}
-func (dropTransport) Close() error                      { return nil }
-
-func failedCount(f *fakeWG) int {
-	f.mu.Lock()
-	defer f.mu.Unlock()
-	return len(f.failed)
-}
-
-func TestManager_InitialTimeoutFailsImmediately(t *testing.T) {
-	wg := newFakeWG()
-	d := NewManager("bbbb", wg, nil) // bbbb > aaaa -> initiator
-	d.Start(dropTransport{})
-	d.retryInterval = 5 * time.Millisecond
-	d.maxRetries = 3
-	defer d.Stop()
-
-	// Bootstrap offer is produced for signalling; no answer ever comes back -> the
-	// initial exchange fails fast.
-	offer, err := d.SignalOffer("aaaa")
-	require.NoError(t, err)
-	require.NotNil(t, offer)
-
-	assert.Eventually(t, func() bool { return failedCount(wg) == 1 }, time.Second, 5*time.Millisecond)
-}
-
-func TestManager_RekeyToleratesKFailures(t *testing.T) {
-	dA, dB, _, wgB, lbB := pair(t)
-	defer dA.Stop()
-	defer dB.Stop()
-
-	// Tighten B's timings before any exchange loop spawns (the loop reads these
-	// fields, so writing them after a loop is running would race).
-	dB.retryInterval = 5 * time.Millisecond
-	dB.maxRetries = 2
-
-	// Establish: bootstrap + data-path-rekeyed so B becomes established and its data
-	// path is usable.
-	bootstrap(t, dA, dB)
-	dA.OnDataPathRekeyed("bbbb", 0)
-	dB.OnDataPathRekeyed("aaaa", 0)
-	require.NotEqual(t, PSK{}, wgB.psk("aaaa"))
-
-	// Drop B's outbound so rekeys can no longer converge.
-	lbB.drop.Store(true)
-
-	// K-1 data-path rekeys must NOT raise OnRekeyFailed.
-	for i := 0; i < DefaultMaxRekeyFailures-1; i++ {
-		_, err := dB.startExchangeTest("aaaa", false, ExchangeID{})
-		require.NoError(t, err)
-		time.Sleep(50 * time.Millisecond)
-	}
-	assert.Equal(t, 0, failedCount(wgB), "no failure before K attempts")
-
-	// The K-th failure raises it once.
-	_, err := dB.startExchangeTest("aaaa", false, ExchangeID{})
-	require.NoError(t, err)
-	assert.Eventually(t, func() bool { return failedCount(wgB) == 1 }, time.Second, 5*time.Millisecond)
-}

client/internal/pqkem/env.go

@@ -1,138 +0,0 @@
-package pqkem
-
-import (
-	"context"
-	"log/slog"
-	"os"
-	"strconv"
-	"strings"
-
-	log "github.com/sirupsen/logrus"
-)
-
-// EnvEnabled is the environment variable that turns the ML-KEM post-quantum
-// exchange on for this client. Accepts on/off aliases plus anything
-// strconv.ParseBool understands (true/false/1/0).
-const EnvEnabled = "NB_ENABLE_PQ_MLKEM"
-
-// Enabled reports whether the ML-KEM PQ exchange is enabled via the environment.
-// An empty or unrecognized value is treated as disabled.
-func Enabled() bool {
-	raw := strings.ToLower(strings.TrimSpace(os.Getenv(EnvEnabled)))
-	switch raw {
-	case "":
-		return false
-	case "on":
-		return true
-	case "off":
-		return false
-	}
-	enabled, err := strconv.ParseBool(raw)
-	if err != nil {
-		log.Warnf("failed to parse %s value %q: %v", EnvEnabled, raw, err)
-		return false
-	}
-	return enabled
-}
-
-// EnvStrict enables strict (fail-closed) mode: block peer traffic until the ML-KEM
-// PSK has been established, instead of the default opportunistic behaviour that lets
-// the tunnel come up classically and upgrades to PQ once the exchange converges.
-const EnvStrict = "NB_PQ_MLKEM_STRICT"
-
-// Strict reports whether strict (fail-closed) mode is enabled via the environment.
-// An empty or unrecognized value is treated as disabled (opportunistic).
-func Strict() bool {
-	switch strings.ToLower(strings.TrimSpace(os.Getenv(EnvStrict))) {
-	case "on":
-		return true
-	case "", "off":
-		return false
-	}
-	enabled, err := strconv.ParseBool(strings.TrimSpace(os.Getenv(EnvStrict)))
-	if err != nil {
-		log.Warnf("failed to parse %s value %q: %v", EnvStrict, os.Getenv(EnvStrict), err)
-		return false
-	}
-	return enabled
-}
-
-// EnvLogLevel overrides the ML-KEM manager's slog level (trace/debug/info/warn/error).
-// Defaults to info. The verbose per-exchange lifecycle logs are emitted at trace.
-const EnvLogLevel = "NB_PQ_MLKEM_LOG_LEVEL"
-
-// LevelTrace is a custom slog level below Debug for the verbose per-exchange lifecycle
-// logs, so they stay off unless NB_PQ_MLKEM_LOG_LEVEL=trace (and the daemon log level
-// is trace, since the records are forwarded to logrus).
-const LevelTrace = slog.LevelDebug - 4
-
-// NewLogger builds the slog logger for the ML-KEM manager. It forwards records to
-// logrus so PQ logs land in the same sink as the rest of the daemon (console +
-// client.log) rather than stdout. Verbosity is gated by EnvLogLevel.
-func NewLogger() *slog.Logger {
-	return slog.New(slogToLogrus{})
-}
-
-func logLevel() slog.Level {
-	switch strings.ToLower(strings.TrimSpace(os.Getenv(EnvLogLevel))) {
-	case "trace":
-		return LevelTrace
-	case "debug":
-		return slog.LevelDebug
-	case "warn":
-		return slog.LevelWarn
-	case "error":
-		return slog.LevelError
-	default:
-		return slog.LevelInfo
-	}
-}
-
-// slogToLogrus is a slog.Handler that forwards records to logrus, so the ML-KEM
-// manager's logs go wherever the daemon's logrus is configured (console + client.log)
-// instead of stdout. Verbosity is gated by EnvLogLevel via logLevel().
-type slogToLogrus struct {
-	fields log.Fields
-}
-
-func (h slogToLogrus) Enabled(_ context.Context, level slog.Level) bool {
-	return level >= logLevel()
-}
-
-func (h slogToLogrus) Handle(_ context.Context, r slog.Record) error {
-	fields := make(log.Fields, len(h.fields)+r.NumAttrs())
-	for k, v := range h.fields {
-		fields[k] = v
-	}
-	r.Attrs(func(a slog.Attr) bool {
-		fields[a.Key] = a.Value.Any()
-		return true
-	})
-	entry := log.WithFields(fields)
-	switch {
-	case r.Level >= slog.LevelError:
-		entry.Error(r.Message)
-	case r.Level >= slog.LevelWarn:
-		entry.Warn(r.Message)
-	case r.Level >= slog.LevelInfo:
-		entry.Info(r.Message)
-	case r.Level >= slog.LevelDebug:
-		entry.Debug(r.Message)
-	default:
-		entry.Trace(r.Message)
-	}
-	return nil
-}
-
-func (h slogToLogrus) WithAttrs(attrs []slog.Attr) slog.Handler {
-	fields := make(log.Fields, len(h.fields)+len(attrs))
-	for k, v := range h.fields {
-		fields[k] = v
-	}
-	for _, a := range attrs {
-		fields[a.Key] = a.Value.Any()
-	}
-	return slogToLogrus{fields: fields}
-}
-
-func (h slogToLogrus) WithGroup(_ string) slog.Handler { return h }

client/internal/pqkem/kem.go

@@ -1,183 +0,0 @@
-// Package pqkem is a spike (NET-1406) for a post-quantum pre-shared-key exchange
-// that could replace Rosenpass. It performs an X25519MLKEM768 hybrid key
-// encapsulation and derives a 32-byte pre-shared key (PSK).
-//
-// The exchange is a single round trip designed to ride the (already
-// authenticated) Signal offer/answer channel:
-//
-//	initiator --Offer(1216B)--> responder
-//	initiator <--Answer(1120B)-- responder
-//
-// Both sides then hold the same PSK, which is bound to the two peers' identities
-// (their peer identity keys) so the derived key cannot be transplanted
-// to a different peer pair even if the transport authentication were bypassed.
-//
-// Combiner note: this follows draft-ietf-tls-ecdhe-mlkem for X25519MLKEM768 — on
-// the wire ML-KEM ‖ X25519 (the draft deliberately reversed the share order for
-// this group), and ML-KEM_ss ‖ X25519_ss as the KDF input. The PSK is derived with
-// HKDF-SHA256 over that hybrid secret, salted with a domain-separation label and
-// bound (via the HKDF info) to the full transcript and the canonicalised peer
-// identities.
-package pqkem
-
-import (
-	"crypto/ecdh"
-	"crypto/hkdf"
-	"crypto/mlkem"
-	"crypto/rand"
-	"crypto/sha256"
-	"fmt"
-)
-
-const (
-	// OfferSize is the initiator message: ML-KEM-768 encapsulation key ‖ X25519 public key
-	// (share order per draft-ietf-tls-ecdhe-mlkem for X25519MLKEM768).
-	OfferSize = mlkem.EncapsulationKeySize768 + 32 // 1216
-	// AnswerSize is the responder message: ML-KEM-768 ciphertext ‖ X25519 public key.
-	AnswerSize = mlkem.CiphertextSize768 + 32 // 1120
-
-	pskLabel = "netbird-pq-psk-v1"
-)
-
-// PSK is the 32-byte derived pre-shared key handed to the consumer to key its channel.
-type PSK [32]byte
-
-// Binding identifies the peer pair the PSK is derived for. Callers set both
-// peer identity keys; the order does not matter (it is canonicalised).
-type Binding struct {
-	LocalID  []byte
-	RemoteID []byte
-}
-
-// Initiator holds the ephemeral secrets between Offer and Finish.
-type Initiator struct {
-	x25519  *ecdh.PrivateKey
-	mlkemDK *mlkem.DecapsulationKey768
-	offer   []byte
-}
-
-// NewInitiator generates the ephemeral X25519 + ML-KEM-768 keypairs.
-func NewInitiator() (*Initiator, error) {
-	x, err := ecdh.X25519().GenerateKey(rand.Reader)
-	if err != nil {
-		return nil, fmt.Errorf("x25519 keygen: %w", err)
-	}
-	dk, err := mlkem.GenerateKey768()
-	if err != nil {
-		return nil, fmt.Errorf("ml-kem keygen: %w", err)
-	}
-
-	offer := make([]byte, 0, OfferSize)
-	offer = append(offer, dk.EncapsulationKey().Bytes()...)
-	offer = append(offer, x.PublicKey().Bytes()...)
-
-	return &Initiator{x25519: x, mlkemDK: dk, offer: offer}, nil
-}
-
-// Offer returns the initiator message to send over Signal.
-func (i *Initiator) Offer() []byte {
-	return i.offer
-}
-
-// Finish consumes the responder's answer and derives the PSK.
-func (i *Initiator) Finish(answer []byte, b Binding) (PSK, error) {
-	if len(answer) != AnswerSize {
-		return PSK{}, fmt.Errorf("answer: got %d bytes, want %d", len(answer), AnswerSize)
-	}
-	ct := answer[:mlkem.CiphertextSize768]
-	peerX := answer[mlkem.CiphertextSize768:]
-
-	ssMLKEM, err := i.mlkemDK.Decapsulate(ct)
-	if err != nil {
-		return PSK{}, fmt.Errorf("ml-kem decapsulate: %w", err)
-	}
-	pub, err := ecdh.X25519().NewPublicKey(peerX)
-	if err != nil {
-		return PSK{}, fmt.Errorf("parse peer x25519: %w", err)
-	}
-	ssX, err := i.x25519.ECDH(pub)
-	if err != nil {
-		return PSK{}, fmt.Errorf("x25519 ecdh: %w", err)
-	}
-
-	return derivePSK(ssMLKEM, ssX, i.offer, answer, b)
-}
-
-// Respond consumes an initiator offer, produces the answer, and derives the PSK.
-func Respond(offer []byte, b Binding) (answer []byte, psk PSK, err error) {
-	if len(offer) != OfferSize {
-		return nil, PSK{}, fmt.Errorf("offer: got %d bytes, want %d", len(offer), OfferSize)
-	}
-	peerEK := offer[:mlkem.EncapsulationKeySize768]
-	peerX := offer[mlkem.EncapsulationKeySize768:]
-
-	ek, err := mlkem.NewEncapsulationKey768(peerEK)
-	if err != nil {
-		return nil, PSK{}, fmt.Errorf("parse peer ml-kem key: %w", err)
-	}
-	ssMLKEM, ct := ek.Encapsulate()
-
-	x, err := ecdh.X25519().GenerateKey(rand.Reader)
-	if err != nil {
-		return nil, PSK{}, fmt.Errorf("x25519 keygen: %w", err)
-	}
-	pub, err := ecdh.X25519().NewPublicKey(peerX)
-	if err != nil {
-		return nil, PSK{}, fmt.Errorf("parse peer x25519: %w", err)
-	}
-	ssX, err := x.ECDH(pub)
-	if err != nil {
-		return nil, PSK{}, fmt.Errorf("x25519 ecdh: %w", err)
-	}
-
-	answer = make([]byte, 0, AnswerSize)
-	answer = append(answer, ct...)
-	answer = append(answer, x.PublicKey().Bytes()...)
-
-	// derivePSK uses the same argument order on both sides; the responder's local
-	// binding is the mirror of the initiator's, canonicalised inside derivePSK.
-	psk, err = derivePSK(ssMLKEM, ssX, offer, answer, b)
-	if err != nil {
-		return nil, PSK{}, err
-	}
-	return answer, psk, nil
-}
-
-// derivePSK runs HKDF-SHA256 over the hybrid shared secret (ML-KEM_ss ‖ X25519_ss,
-// per draft-ietf-tls-ecdhe-mlkem), salted with the domain-separation label, and binds
-// the result — via the HKDF info — to the full transcript (offer ‖ answer) and the
-// canonicalised peer identities, so the PSK cannot be transplanted to another peer
-// pair or a different exchange.
-func derivePSK(ssMLKEM, ssX, offer, answer []byte, b Binding) (PSK, error) {
-	// A PSK not bound to both peer identities could be transplanted to a different peer
-	// pair, so refuse to derive one from an empty binding.
-	if len(b.LocalID) == 0 || len(b.RemoteID) == 0 {
-		return PSK{}, fmt.Errorf("empty peer identity binding")
-	}
-	lo, hi := canonicalPair(b.LocalID, b.RemoteID)
-
-	ikm := make([]byte, 0, len(ssMLKEM)+len(ssX))
-	ikm = append(ikm, ssMLKEM...)
-	ikm = append(ikm, ssX...)
-
-	info := make([]byte, 0, len(offer)+len(answer)+len(lo)+len(hi))
-	info = append(info, offer...)
-	info = append(info, answer...)
-	info = append(info, lo...)
-	info = append(info, hi...)
-
-	var psk PSK
-	key, err := hkdf.Key(sha256.New, ikm, []byte(pskLabel), string(info), len(psk))
-	if err != nil {
-		return PSK{}, fmt.Errorf("hkdf derive psk: %w", err)
-	}
-	copy(psk[:], key)
-	return psk, nil
-}
-
-func canonicalPair(a, b []byte) (lo, hi []byte) {
-	if string(a) <= string(b) {
-		return a, b
-	}
-	return b, a
-}

client/internal/pqkem/kem_framing_test.go

@@ -1,105 +0,0 @@
-package pqkem
-
-import (
-	"crypto/mlkem"
-	"testing"
-
-	"github.com/stretchr/testify/require"
-)
-
-// TestExchange_TamperedCiphertextFailsClosed verifies the core fail-closed
-// property: mutating the ML-KEM ciphertext in the answer does not error (ML-KEM
-// uses implicit rejection — Decapsulate always returns a value) but yields a
-// different shared secret, so the initiator derives a PSK that does NOT match the
-// responder's. A mismatched PSK means WireGuard passes no bytes: tamper => no data.
-func TestExchange_TamperedCiphertextFailsClosed(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-
-	answer, pskB, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-
-	tampered := append([]byte(nil), answer...)
-	tampered[0] ^= 0xff // flip a bit in the ML-KEM ciphertext
-
-	pskA, err := init.Finish(tampered, Binding{LocalID: wgA, RemoteID: wgB})
-	require.NoError(t, err, "implicit rejection: decapsulate still succeeds")
-	require.NotEqual(t, pskB, pskA, "tampered ciphertext must not yield the responder's PSK")
-}
-
-// TestExchange_TamperedX25519ShareDiverges flips a byte in the answer's X25519
-// share: the classical half of the hybrid secret changes, so the derived PSK
-// diverges from the responder's (fail-closed on the ECDH half too).
-func TestExchange_TamperedX25519ShareDiverges(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-
-	answer, pskB, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-
-	tampered := append([]byte(nil), answer...)
-	tampered[mlkem.CiphertextSize768] ^= 0x01 // first byte of the X25519 public key
-
-	pskA, err := init.Finish(tampered, Binding{LocalID: wgA, RemoteID: wgB})
-	// Either the point is rejected (error) or the ECDH differs (different PSK);
-	// in both cases the honest PSK is never reproduced.
-	if err == nil {
-		require.NotEqual(t, pskB, pskA, "tampered X25519 share must not yield the responder's PSK")
-	}
-}
-
-// TestExchange_AllZeroX25519Rejected feeds an all-zero X25519 share (a low-order
-// point) in the answer. The stdlib ECDH must reject it, so Finish errors rather
-// than deriving a PSK from a degenerate secret.
-func TestExchange_AllZeroX25519Rejected(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-
-	answer, _, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-
-	bad := append([]byte(nil), answer...)
-	for i := mlkem.CiphertextSize768; i < len(bad); i++ {
-		bad[i] = 0
-	}
-
-	_, err = init.Finish(bad, Binding{LocalID: wgA, RemoteID: wgB})
-	require.Error(t, err, "all-zero X25519 share (low-order point) must be rejected")
-}
-
-// TestExchange_SizeBoundaries locks the exact-length framing checks: one byte
-// short or long on either message is rejected, not silently truncated/padded.
-func TestExchange_SizeBoundaries(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-	offer := init.Offer()
-
-	_, _, err = Respond(offer[:OfferSize-1], Binding{})
-	require.Error(t, err, "offer one byte short")
-	_, _, err = Respond(append(append([]byte(nil), offer...), 0), Binding{})
-	require.Error(t, err, "offer one byte long")
-
-	answer, _, err := Respond(offer, Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-
-	_, err = init.Finish(answer[:AnswerSize-1], Binding{LocalID: wgA, RemoteID: wgB})
-	require.Error(t, err, "answer one byte short")
-	_, err = init.Finish(append(append([]byte(nil), answer...), 0), Binding{LocalID: wgA, RemoteID: wgB})
-	require.Error(t, err, "answer one byte long")
-}
-
-// TestExchange_BindingIsSymmetric confirms the canonicalisation: the two peers
-// pass their identities in opposite (Local, Remote) order yet derive the same PSK,
-// so identity binding does not depend on who is initiator vs responder.
-func TestExchange_BindingIsSymmetric(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-
-	answer, pskB, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-
-	pskA, err := init.Finish(answer, Binding{LocalID: wgA, RemoteID: wgB})
-	require.NoError(t, err)
-
-	require.Equal(t, pskB, pskA, "swapped Local/Remote order must canonicalise to the same PSK")
-}

client/internal/pqkem/kem_test.go

@@ -1,106 +0,0 @@
-package pqkem
-
-import (
-	"testing"
-	"time"
-
-	"github.com/stretchr/testify/require"
-)
-
-var (
-	wgA = []byte("peer-A-wireguard-pubkey-32bytes!")
-	wgB = []byte("peer-B-wireguard-pubkey-32bytes!")
-)
-
-func TestExchange_DerivesMatchingPSK(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-
-	require.Len(t, init.Offer(), OfferSize)
-
-	answer, pskB, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-	require.Len(t, answer, AnswerSize)
-
-	pskA, err := init.Finish(answer, Binding{LocalID: wgA, RemoteID: wgB})
-	require.NoError(t, err)
-
-	require.Equal(t, pskB, pskA, "both sides must derive the same PSK")
-	require.NotEqual(t, PSK{}, pskA, "PSK must not be zero")
-}
-
-func TestExchange_PSKBoundToPeerIdentities(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-
-	answer, _, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-
-	// Finish twice over the SAME KEM material (same offer/answer/secrets), changing only
-	// the peer identity binding: the differing PSK is attributable to the binding alone.
-	pskHonest, err := init.Finish(answer, Binding{LocalID: wgA, RemoteID: wgB})
-	require.NoError(t, err)
-
-	wgC := []byte("peer-C-wireguard-pubkey-32bytes!")
-	pskWrong, err := init.Finish(answer, Binding{LocalID: wgA, RemoteID: wgC})
-	require.NoError(t, err)
-
-	require.NotEqual(t, pskHonest, pskWrong, "PSK must be bound to the peer pair")
-}
-
-func TestExchange_RejectsEmptyBinding(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-	answer, _, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-
-	// A PSK not bound to both identities could be transplanted to another peer pair.
-	_, err = init.Finish(answer, Binding{})
-	require.Error(t, err, "empty binding must be rejected")
-	_, err = init.Finish(answer, Binding{LocalID: wgA})
-	require.Error(t, err, "missing RemoteID must be rejected")
-	_, _, err = Respond(init.Offer(), Binding{RemoteID: wgA})
-	require.Error(t, err, "missing LocalID must be rejected")
-}
-
-func TestExchange_RejectsMalformedMessages(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-
-	_, _, err = Respond(init.Offer()[:10], Binding{})
-	require.Error(t, err)
-
-	_, err = init.Finish([]byte("too short"), Binding{})
-	require.Error(t, err)
-}
-
-// TestExchange_ReportSizesAndTiming is a spike measurement, not a pass/fail gate.
-// Run with: go test -run TestExchange_ReportSizesAndTiming -v ./client/internal/pqkem/
-func TestExchange_ReportSizesAndTiming(t *testing.T) {
-	const iters = 200
-
-	var tInit, tResp, tFinish time.Duration
-	for i := 0; i < iters; i++ {
-		s0 := time.Now()
-		init, err := NewInitiator()
-		require.NoError(t, err)
-		tInit += time.Since(s0)
-
-		s1 := time.Now()
-		answer, _, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-		require.NoError(t, err)
-		tResp += time.Since(s1)
-
-		s2 := time.Now()
-		_, err = init.Finish(answer, Binding{LocalID: wgA, RemoteID: wgB})
-		require.NoError(t, err)
-		tFinish += time.Since(s2)
-	}
-
-	t.Logf("wire sizes:   offer=%d B   answer=%d B   (Rosenpass static pubkey ~524160 B)", OfferSize, AnswerSize)
-	t.Logf("total on-wire per handshake: %d B  (~%.0fx smaller than RP static key)", OfferSize+AnswerSize, 524160.0/float64(OfferSize+AnswerSize))
-	t.Logf("avg NewInitiator (keygen): %s", tInit/iters)
-	t.Logf("avg Respond (encaps+dh):   %s", tResp/iters)
-	t.Logf("avg Finish (decaps+dh):    %s", tFinish/iters)
-	t.Logf("avg full handshake CPU:    %s", (tInit+tResp+tFinish)/iters)
-}

client/internal/pqkem/manager.go

@@ -1,467 +0,0 @@
-package pqkem
-
-import (
-	"context"
-	"crypto/rand"
-	"fmt"
-	"log/slog"
-	"net/netip"
-	"sync"
-	"time"
-)
-
-const (
-	// DefaultRetryInterval is how often the initiator retransmits its outstanding
-	// data-path offer while awaiting the answer.
-	DefaultRetryInterval = 2 * time.Second
-	// DefaultMaxRetries bounds how many ticks an exchange may run before it is
-	// declared failed. The convergence deadline is thus MaxRetries * RetryInterval.
-	DefaultMaxRetries = 10
-	// DefaultMaxRekeyFailures is how many consecutive rekey (non-initial) failures
-	// are tolerated before OnRekeyFailed. The initial exchange fails immediately.
-	DefaultMaxRekeyFailures = 3
-
-	// rotationActivityWindow gates rotation on recent real-data activity: a rekey
-	// clocks a rotation only if the peer exchanged user data within this window. It
-	// must stay shorter than the data path's rekey interval (WireGuard
-	// REKEY_AFTER_TIME ~120s) so the rotation's own traffic — which itself renews the
-	// activity signal — ages out before the next rekey, letting an idle tunnel stop
-	// rotating instead of self-sustaining.
-	rotationActivityWindow = 90 * time.Second
-)
-
-// LocalID and RemoteID are peer identity keys (e.g. WireGuard public keys). They are
-// distinct types so the local and a remote identity cannot be mixed up.
-type (
-	LocalID  string
-	RemoteID string
-)
-
-// Transport is the data-path socket the Manager drives (the analogue of
-// go-rosenpass's Conn). It is a dumb mover of bytes to/from endpoints: the Manager
-// owns the remoteID<->endpoint routing and hands the transport a resolved endpoint
-// to Send, and reverse-resolves the source of each inbound datagram. Its lifecycle
-// belongs to the Manager (Run at Start, Close at Stop).
-type Transport interface {
-	// Send delivers msg to the given data-path endpoint.
-	Send(endpoint netip.AddrPort, msg []byte) error
-	// LocalPort is the bound local UDP port, announced to peers so they know where
-	// to send data-path messages.
-	LocalPort() int
-	// Run starts delivering inbound datagrams as (source endpoint, msg) to onInbound
-	// and returns immediately; it runs until Close.
-	Run(onInbound func(src netip.AddrPort, msg []byte))
-	// Close stops delivery and releases the socket.
-	Close() error
-}
-
-// exchangeState is the single source of truth for an exchange's role and phase.
-type exchangeState uint8
-
-const (
-	stateReserved       exchangeState = iota // responder: deriving the answer
-	stateAwaitingAnswer                      // initiator: offer sent, awaiting the answer
-	stateAwaitingRekey                       // initiator: PSK derived+set, awaiting OnDataPathRekeyed to chain the next offer
-	stateAwaitingAck                         // responder: answer sent, awaiting the next offer that acks this exchange
-)
-
-// exchangeCtl holds all state for one in-flight exchange with a peer, under the
-// Manager's single lock. state drives every decision. lastSent is the current
-// data-path retransmit payload (the offer, for the initiator). initiator is the
-// ephemeral handle used at Finish; pendingPSK is the responder's derived key.
-// viaSignal records that the offer went to the host for the signalling channel, so
-// the loop does not retransmit it on the data path. Only the initiator runs a
-// retransmit loop, so only it sets cancel.
-type exchangeCtl struct {
-	id         ExchangeID
-	state      exchangeState
-	startedAt  time.Time
-	cancel     context.CancelFunc
-	lastSent   []byte
-	initiator  *Initiator
-	pendingPSK PSK
-	viaSignal  bool
-}
-
-// Manager is the stateful orchestrator — the analogue of go-rosenpass's Server. It
-// drives the X25519MLKEM768 exchange, owns the peer endpoint routing and the data-path
-// transport, and surfaces the derived PSK and convergence to the host via
-// CallbackHandler. It is event-driven: the bootstrap is triggered by the host
-// (SignalOffer) and each rotation is clocked by OnDataPathRekeyed. The cryptography is
-// the pure kem.go primitives; all state lives here under one lock.
-type Manager struct {
-	localID   LocalID
-	cbHandler CallbackHandler
-	logger    *slog.Logger
-
-	retryInterval    time.Duration
-	maxRetries       int
-	maxRekeyFailures int
-
-	rootCtx    context.Context
-	rootCancel context.CancelFunc
-
-	mu          sync.Mutex
-	transport   Transport
-	exchanges   map[RemoteID]*exchangeCtl   // in-flight exchange per peer
-	established map[RemoteID]bool           // peer has completed at least one exchange
-	failures    map[RemoteID]int            // consecutive rekey failures per peer
-	psks        map[RemoteID]PSK            // latest derived PSK per peer (pulled at WG peer-config time)
-	capable     map[RemoteID]bool           // peer runs the KEM (advertised a PQ port); false = known non-capable
-	peerAddrs   map[RemoteID]netip.AddrPort // remoteID -> data-path endpoint (send routing)
-	peersByAddr map[netip.AddrPort]RemoteID // reverse: source endpoint -> remoteID (inbound)
-	wait        sync.WaitGroup
-}
-
-// NewManager builds a manager for the local peer identified by its peer identity key
-// (used for the deterministic initiator role and the identity binding). A nil logger
-// falls back to slog.Default(). Install the data-path transport with Start.
-func NewManager(localID LocalID, h CallbackHandler, logger *slog.Logger) *Manager {
-	if logger == nil {
-		logger = slog.Default()
-	}
-	ctx, cancel := context.WithCancel(context.Background())
-	return &Manager{
-		localID:          localID,
-		cbHandler:        h,
-		logger:           logger,
-		retryInterval:    DefaultRetryInterval,
-		maxRetries:       DefaultMaxRetries,
-		maxRekeyFailures: DefaultMaxRekeyFailures,
-		rootCtx:          ctx,
-		rootCancel:       cancel,
-		exchanges:        make(map[RemoteID]*exchangeCtl),
-		established:      make(map[RemoteID]bool),
-		failures:         make(map[RemoteID]int),
-		psks:             make(map[RemoteID]PSK),
-		capable:          make(map[RemoteID]bool),
-		peerAddrs:        make(map[RemoteID]netip.AddrPort),
-		peersByAddr:      make(map[netip.AddrPort]RemoteID),
-	}
-}
-
-// Start installs the data-path transport and begins its inbound delivery. The Manager
-// owns it from here; Stop closes it.
-func (m *Manager) Start(t Transport) {
-	m.mu.Lock()
-	m.transport = t
-	m.mu.Unlock()
-	if t != nil {
-		t.Run(m.onDataPathInbound)
-	}
-}
-
-// LocalPort is the data-path transport's bound UDP port (0 if no transport), to be
-// announced to peers.
-func (m *Manager) LocalPort() int {
-	m.mu.Lock()
-	t := m.transport
-	m.mu.Unlock()
-	if t == nil {
-		return 0
-	}
-	return t.LocalPort()
-}
-
-// IsInitiator reports whether the local peer drives the exchange for this remote
-// peer. Roles are deterministic (lexicographic identity-key compare) so exactly one
-// side initiates, mirroring how Rosenpass picks its handshake initiator.
-func (m *Manager) IsInitiator(remoteID RemoteID) bool {
-	return string(m.localID) > string(remoteID)
-}
-
-// PSK returns the latest PSK derived for the peer, for the host to program at WG
-// peer-config time (the pull path). ok is false until an exchange has derived one.
-func (m *Manager) PSK(remoteID RemoteID) (PSK, bool) {
-	m.mu.Lock()
-	defer m.mu.Unlock()
-	psk, ok := m.psks[remoteID]
-	return psk, ok
-}
-
-// trace logs at LevelTrace, the verbose per-exchange lifecycle level gated by
-// NB_PQ_MLKEM_LOG_LEVEL=trace.
-func (m *Manager) trace(msg string, args ...any) {
-	m.logger.Log(context.Background(), LevelTrace, msg, args...)
-}
-
-// AddPeer registers where a peer's data-path messages are sent and received: its
-// overlay endpoint (IP:port). This is pure routing and says nothing about capability —
-// PQ capability is decided solely from the peer's KEM payload (see processOffer /
-// processAnswer / MarkNonCapable), never from an endpoint or port.
-func (m *Manager) AddPeer(remoteID RemoteID, endpoint netip.AddrPort) {
-	if !endpoint.IsValid() {
-		return
-	}
-	m.mu.Lock()
-	if old, ok := m.peerAddrs[remoteID]; ok {
-		delete(m.peersByAddr, old)
-	}
-	m.peerAddrs[remoteID] = endpoint
-	m.peersByAddr[endpoint] = remoteID
-	m.mu.Unlock()
-}
-
-// MarkNonCapable records that a peer does not run the KEM: it answered our offer with
-// no KEM material over signalling (the capability signal is the peer's payload, not its
-// optional data-path port). Any in-flight exchange is cancelled and further offers are
-// suppressed (see SignalOffer), so a non-PQ peer never drives the rekey-recovery storm.
-// An already-established peer is left untouched — a stray empty answer must not tear
-// down a working PQ session.
-func (m *Manager) MarkNonCapable(remoteID RemoteID) {
-	m.mu.Lock()
-	defer m.mu.Unlock()
-	if m.established[remoteID] {
-		return
-	}
-	if prev, ok := m.capable[remoteID]; ok && !prev {
-		return // already known non-capable, nothing to do
-	}
-	m.capable[remoteID] = false
-	if ex := m.exchanges[remoteID]; ex != nil {
-		if ex.cancel != nil {
-			ex.cancel()
-		}
-		delete(m.exchanges, remoteID)
-	}
-	m.trace("pqkem: peer advertises no PQ service — treating as non-capable, no KEM attempted", "peer", remoteID)
-}
-
-// RemovePeer stops any in-flight exchange for a peer and drops its state and routing.
-func (m *Manager) RemovePeer(remoteID RemoteID) {
-	m.mu.Lock()
-	if ex, ok := m.exchanges[remoteID]; ok {
-		if ex.cancel != nil {
-			ex.cancel()
-		}
-		delete(m.exchanges, remoteID)
-	}
-	delete(m.established, remoteID)
-	delete(m.failures, remoteID)
-	delete(m.psks, remoteID)
-	delete(m.capable, remoteID)
-	if ep, ok := m.peerAddrs[remoteID]; ok {
-		delete(m.peersByAddr, ep)
-		delete(m.peerAddrs, remoteID)
-	}
-	m.mu.Unlock()
-}
-
-// Stop cancels all in-flight exchanges, closes the transport, and waits for the
-// exchange goroutines to exit.
-func (m *Manager) Stop() {
-	// Cancel the root context under the lock, before Wait: startExchangeLocked checks
-	// rootCtx.Err() under the same lock before it Adds to the wait group, so once Stop
-	// has cancelled here no new Add can race Wait.
-	m.mu.Lock()
-	m.rootCancel()
-	m.mu.Unlock()
-	m.wait.Wait()
-	m.mu.Lock()
-	t := m.transport
-	m.transport = nil
-	m.exchanges = make(map[RemoteID]*exchangeCtl)
-	m.psks = make(map[RemoteID]PSK)
-	m.mu.Unlock()
-	if t != nil {
-		if err := t.Close(); err != nil {
-			m.logger.Warn("pqkem: closing data-path transport", "err", err)
-		}
-	}
-}
-
-// ---- Signalling channel (host-driven; rides the host's negotiation) ----
-
-// SignalOffer returns the KEM offer for the host to embed in its outgoing offer to
-// remoteID (bootstrap). It returns (nil, nil) when the local peer is not the
-// initiator. It is idempotent for an in-flight bootstrap: a repeat call returns the
-// same offer rather than starting a new exchange.
-//
-// A signal re-negotiation always re-bootstraps (fresh exchange): the remote may have
-// restarted and lost its PSK, so reusing a locally frozen one would desync. The derived
-// PSK still survives idle in the manager (dropped only on account-level peer removal),
-// so a pure lazy wake with no re-negotiation reuses it via the conn's WG-config pull.
-func (m *Manager) SignalOffer(remoteID RemoteID) ([]byte, error) {
-	if !m.IsInitiator(remoteID) {
-		return nil, nil
-	}
-	m.mu.Lock()
-	if capable, ok := m.capable[remoteID]; ok && !capable {
-		m.mu.Unlock()
-		return nil, nil // peer does not run the KEM; do not offer (avoids a failure/reoffer loop)
-	}
-	// Idempotent while a signalling bootstrap is in flight OR already derived a PSK but
-	// not yet chained a rotation (awaitingRekey): return the SAME offer instead of
-	// starting a new exchange. This matters when the controller both offers on its own
-	// guard AND re-offers in response to the responder's offer — without this, the
-	// second call would start a fresh exchange (a different PSK) and desync the peers.
-	if ex := m.exchanges[remoteID]; ex != nil && ex.viaSignal &&
-		(ex.state == stateAwaitingAnswer || ex.state == stateAwaitingRekey) {
-		last := ex.lastSent
-		m.mu.Unlock()
-		return last, nil
-	}
-	// Hold the lock across the check above and the install so a concurrent SignalOffer
-	// for the same peer can't also start an exchange. bootstrap offer acks nothing.
-	raw, err := m.startExchangeLocked(remoteID, true, ExchangeID{})
-	m.mu.Unlock()
-	return raw, err
-}
-
-// ShouldSendBootstrapOffer reports whether we should emit a fresh KEM offer to kick a
-// bootstrap for this peer. True only if we are the initiator, the peer is not known
-// non-capable, and no exchange is already in flight. The host uses this when it (as the
-// controller) receives the responder's offer: it replies with a KEM offer exactly once
-// to start the exchange, and ignores further responder offers while one is in flight,
-// avoiding an offer-per-offer runaway.
-func (m *Manager) ShouldSendBootstrapOffer(remoteID RemoteID) bool {
-	if !m.IsInitiator(remoteID) {
-		return false
-	}
-	m.mu.Lock()
-	defer m.mu.Unlock()
-	if capable, ok := m.capable[remoteID]; ok && !capable {
-		return false
-	}
-	return m.exchanges[remoteID] == nil
-}
-
-// SignalOnOffer processes a KEM offer the host extracted from an incoming offer and
-// returns the KEM answer for the host to embed in its outgoing answer.
-func (m *Manager) SignalOnOffer(remoteID RemoteID, offer []byte) ([]byte, error) {
-	typ, msg, err := Decode(offer)
-	if err != nil {
-		return nil, fmt.Errorf("decode signal offer from %s: %w", remoteID, err)
-	}
-	if typ != MsgOffer {
-		return nil, fmt.Errorf("expected offer from %s, got type %d", remoteID, typ)
-	}
-	return m.processOffer(remoteID, msg.(*OfferMsg))
-}
-
-// SignalOnAnswer processes a KEM answer the host extracted from an incoming answer.
-// There is no reply: the next offer (over the data path) acknowledges this exchange.
-func (m *Manager) SignalOnAnswer(remoteID RemoteID, answer []byte) error {
-	typ, msg, err := Decode(answer)
-	if err != nil {
-		return fmt.Errorf("decode signal answer from %s: %w", remoteID, err)
-	}
-	if typ != MsgAnswer {
-		return fmt.Errorf("expected answer from %s, got type %d", remoteID, typ)
-	}
-	return m.processAnswer(remoteID, msg.(*AnswerMsg))
-}
-
-// ---- Data path ----
-
-// onDataPathInbound is the transport's inbound handler: it reverse-resolves the
-// source endpoint to a peer and dispatches. Unknown sources are dropped.
-func (m *Manager) onDataPathInbound(src netip.AddrPort, msg []byte) {
-	m.mu.Lock()
-	remoteID, ok := m.peersByAddr[src]
-	m.mu.Unlock()
-	if !ok {
-		return
-	}
-	if err := m.OnDataPathMessage(remoteID, msg); err != nil {
-		m.trace("pqkem: inbound", "peer", remoteID, "err", err)
-	}
-}
-
-// OnDataPathMessage handles a KEM message received over the data path from remoteID
-// and pushes any reply back over the data path.
-func (m *Manager) OnDataPathMessage(remoteID RemoteID, raw []byte) error {
-	typ, msg, err := Decode(raw)
-	if err != nil {
-		return fmt.Errorf("decode data-path msg from %s: %w", remoteID, err)
-	}
-	switch typ {
-	case MsgOffer:
-		answer, err := m.processOffer(remoteID, msg.(*OfferMsg))
-		if err != nil {
-			return err
-		}
-		if answer == nil {
-			return nil
-		}
-		return m.pushDataPath(remoteID, answer)
-	case MsgAnswer:
-		return m.processAnswer(remoteID, msg.(*AnswerMsg))
-	default:
-		return fmt.Errorf("unhandled data-path message type %d from %s", typ, remoteID)
-	}
-}
-
-// OnDataPathRekeyed clocks the next chained PSK rotation on a fresh data-path rekey
-// (fired on first establishment AND every rekey). If we are the initiator that just
-// derived a PSK, it chains the next exchange: a fresh offer over the data path that
-// acknowledges the just-completed one (its arrival under the new key proves to the
-// responder the key works). sinceActivity is how long ago the peer last exchanged real
-// user data; past rotationActivityWindow the tunnel is treated as idle and rotation is
-// skipped — an idle tunnel has nothing to protect, and rotating would emit data-path
-// traffic that keeps the peer artificially active (see conn.onWGCheckSuccess).
-func (m *Manager) OnDataPathRekeyed(remoteID RemoteID, sinceActivity time.Duration) {
-	if sinceActivity >= rotationActivityWindow {
-		m.trace("pqkem: peer idle, skipping data-path rotation", "peer", remoteID, "since_activity", sinceActivity)
-		return
-	}
-
-	m.mu.Lock()
-	ex := m.exchanges[remoteID]
-	chain := ex != nil && ex.state == stateAwaitingRekey
-	if !chain {
-		m.mu.Unlock()
-		m.trace("pqkem: data-path rekey signal", "peer", remoteID, "chaining", false)
-		return
-	}
-	// Hold the lock across the awaitingRekey check and the install so two rekey clocks
-	// can't each start a chained exchange for the same peer.
-	offer, err := m.startExchangeLocked(remoteID, false, ex.id)
-	m.mu.Unlock()
-
-	m.trace("pqkem: data-path rekey signal", "peer", remoteID, "chaining", true)
-	if err != nil {
-		m.logger.Error("pqkem: chain offer failed to start", "peer", remoteID, "err", err)
-		return
-	}
-	if err := m.pushDataPath(remoteID, offer); err != nil {
-		m.logger.Warn("pqkem: send chain offer failed", "peer", remoteID, "err", err)
-		return
-	}
-	m.trace("pqkem: chain offer sent over data path", "peer", remoteID)
-}
-
-// OnDataPathDown notifies that the peer's data path went down. Rotations resume once
-// the host re-bootstraps over signalling on reconnect; in-flight data-path sends will
-// simply fail until then. Reserved as an explicit hook.
-func (m *Manager) OnDataPathDown(remoteID RemoteID) {}
-
-// ---- internals ----
-
-// pushDataPath resolves the peer's endpoint and sends over the data-path transport,
-// erroring if the peer is unknown or no transport is set.
-func (m *Manager) pushDataPath(remoteID RemoteID, msg []byte) error {
-	m.mu.Lock()
-	ep, ok := m.peerAddrs[remoteID]
-	t := m.transport
-	m.mu.Unlock()
-	if !ok {
-		return fmt.Errorf("no data-path endpoint for peer %s", remoteID)
-	}
-	if t == nil {
-		return fmt.Errorf("no data-path transport")
-	}
-	return t.Send(ep, msg)
-}
-
-func (m *Manager) binding(remoteID RemoteID) Binding {
-	return Binding{LocalID: []byte(m.localID), RemoteID: []byte(remoteID)}
-}
-
-func newExchangeID() (ExchangeID, error) {
-	var id ExchangeID
-	if _, err := rand.Read(id[:]); err != nil {
-		return ExchangeID{}, fmt.Errorf("generate exchange id: %w", err)
-	}
-	return id, nil
-}

client/internal/pqkem/manager_test.go

@@ -1,201 +0,0 @@
-package pqkem
-
-import (
-	"fmt"
-	"net/netip"
-	"sync"
-	"sync/atomic"
-	"testing"
-
-	"github.com/stretchr/testify/require"
-)
-
-// netSwitch is an in-memory UDP fabric: transports register their endpoint and get
-// datagrams delivered to their inbound handler.
-type netSwitch struct {
-	mu sync.Mutex
-	h  map[netip.AddrPort]func(netip.AddrPort, []byte)
-}
-
-func newSwitch() *netSwitch {
-	return &netSwitch{h: map[netip.AddrPort]func(netip.AddrPort, []byte){}}
-}
-
-func (s *netSwitch) register(ep netip.AddrPort, fn func(netip.AddrPort, []byte)) {
-	s.mu.Lock()
-	s.h[ep] = fn
-	s.mu.Unlock()
-}
-
-func (s *netSwitch) deliver(dst, src netip.AddrPort, msg []byte) error {
-	s.mu.Lock()
-	fn := s.h[dst]
-	s.mu.Unlock()
-	if fn == nil {
-		return fmt.Errorf("no route to %s", dst)
-	}
-	fn(src, msg)
-	return nil
-}
-
-// loopback is an endpoint-based pqkem.Transport over a netSwitch, with a switchable
-// drop flag.
-type loopback struct {
-	ep   netip.AddrPort
-	sw   *netSwitch
-	drop atomic.Bool
-}
-
-func (l *loopback) Send(dst netip.AddrPort, msg []byte) error {
-	if l.drop.Load() {
-		return nil
-	}
-	return l.sw.deliver(dst, l.ep, append([]byte(nil), msg...))
-}
-
-func (l *loopback) LocalPort() int                             { return int(l.ep.Port()) }
-func (l *loopback) Run(onInbound func(netip.AddrPort, []byte)) { l.sw.register(l.ep, onInbound) }
-func (l *loopback) Close() error                               { return nil }
-
-type fakeWG struct {
-	mu     sync.Mutex
-	psks   map[RemoteID]PSK
-	failed []RemoteID
-}
-
-func newFakeWG() *fakeWG { return &fakeWG{psks: map[RemoteID]PSK{}} }
-
-// startExchangeTest drives startExchangeLocked with the lock held, for tests that kick an
-// exchange directly (production callers hold m.mu across their idempotency check).
-func (m *Manager) startExchangeTest(remoteID RemoteID, viaSignal bool, ackID ExchangeID) ([]byte, error) {
-	m.mu.Lock()
-	defer m.mu.Unlock()
-	return m.startExchangeLocked(remoteID, viaSignal, ackID)
-}
-
-func (f *fakeWG) OnNewPSKReady(remoteID RemoteID, psk PSK) error {
-	f.mu.Lock()
-	defer f.mu.Unlock()
-	f.psks[remoteID] = psk
-	return nil
-}
-
-func (f *fakeWG) OnRekeyFailed(remoteID RemoteID) error {
-	f.mu.Lock()
-	defer f.mu.Unlock()
-	f.failed = append(f.failed, remoteID)
-	return nil
-}
-
-func (f *fakeWG) psk(peer RemoteID) PSK {
-	f.mu.Lock()
-	defer f.mu.Unlock()
-	return f.psks[peer]
-}
-
-var (
-	epA = netip.MustParseAddrPort("100.64.0.1:51833")
-	epB = netip.MustParseAddrPort("100.64.0.2:51833")
-)
-
-// pair builds two wired managers (B is the initiator, "bbbb" > "aaaa") sharing a
-// netSwitch, with each peer's data-path endpoint registered. lbB is B's loopback
-// (for toggling drop).
-func pair(t *testing.T) (dA, dB *Manager, wgA, wgB *fakeWG, lbB *loopback) {
-	t.Helper()
-	sw := newSwitch()
-	wgA = newFakeWG()
-	wgB = newFakeWG()
-	dA = NewManager("aaaa", wgA, nil)
-	dB = NewManager("bbbb", wgB, nil)
-	dA.Start(&loopback{ep: epA, sw: sw})
-	lbB = &loopback{ep: epB, sw: sw}
-	dB.Start(lbB)
-	dA.AddPeer("bbbb", epB)
-	dB.AddPeer("aaaa", epA)
-	return dA, dB, wgA, wgB, lbB
-}
-
-// bootstrap runs the signalling offer/answer (the test plays the host carrying bytes).
-func bootstrap(t *testing.T, dA, dB *Manager) {
-	t.Helper()
-	offer, err := dB.SignalOffer("aaaa")
-	require.NoError(t, err)
-	require.NotNil(t, offer)
-	answer, err := dA.SignalOnOffer("bbbb", offer)
-	require.NoError(t, err)
-	require.NotNil(t, answer)
-	require.NoError(t, dB.SignalOnAnswer("aaaa", answer))
-}
-
-func TestManager_BootstrapDerivesSamePSK(t *testing.T) {
-	dA, dB, wgA, wgB, _ := pair(t)
-	defer dA.Stop()
-	defer dB.Stop()
-
-	bootstrap(t, dA, dB)
-
-	pskA := wgA.psk("bbbb")
-	pskB := wgB.psk("aaaa")
-	require.NotEqual(t, PSK{}, pskA)
-	require.Equal(t, pskB, pskA, "both sides derive the same PSK from the bootstrap exchange")
-}
-
-func TestManager_ChainRotatesAndAcks(t *testing.T) {
-	dA, dB, wgA, wgB, _ := pair(t)
-	defer dA.Stop()
-	defer dB.Stop()
-
-	bootstrap(t, dA, dB)
-	psk1 := wgB.psk("aaaa")
-
-	// Data path up: B (initiator) chains the next offer over the data path, which
-	// rotates both to a fresh PSK and acknowledges A.
-	dA.OnDataPathRekeyed("bbbb", 0)
-	dB.OnDataPathRekeyed("aaaa", 0)
-
-	psk2A := wgA.psk("bbbb")
-	psk2B := wgB.psk("aaaa")
-	require.Equal(t, psk2B, psk2A, "both sides converge on the rotated PSK")
-	require.NotEqual(t, psk1, psk2B, "the chain rotated to a new PSK")
-}
-
-func TestManager_RotationSkippedWhenIdle(t *testing.T) {
-	dA, dB, wgA, wgB, _ := pair(t)
-	defer dA.Stop()
-	defer dB.Stop()
-
-	bootstrap(t, dA, dB)
-	psk1 := wgB.psk("aaaa")
-	require.NotEqual(t, PSK{}, psk1)
-
-	// Idle: the peer's last real-data activity is older than the window, so a rekey
-	// must NOT clock a rotation.
-	dA.OnDataPathRekeyed("bbbb", rotationActivityWindow)
-	dB.OnDataPathRekeyed("aaaa", rotationActivityWindow)
-	require.Equal(t, psk1, wgB.psk("aaaa"), "idle peer must not rotate the PSK")
-	require.Equal(t, psk1, wgA.psk("bbbb"), "idle peer must not rotate the PSK")
-
-	// Active: activity within the window clocks the rotation as usual.
-	dA.OnDataPathRekeyed("bbbb", rotationActivityWindow-1)
-	dB.OnDataPathRekeyed("aaaa", rotationActivityWindow-1)
-	psk2 := wgB.psk("aaaa")
-	require.NotEqual(t, psk1, psk2, "recent activity must clock a rotation")
-	require.Equal(t, psk2, wgA.psk("bbbb"), "both sides converge on the rotated PSK")
-}
-
-func TestManager_NonInitiatorReturnsNoOffer(t *testing.T) {
-	dA := NewManager("aaaa", newFakeWG(), nil)
-	defer dA.Stop()
-
-	offer, err := dA.SignalOffer("bbbb") // not the initiator vs "bbbb"
-	require.NoError(t, err)
-	require.Nil(t, offer)
-}
-
-func TestManager_StopIsIdempotent(t *testing.T) {
-	dA := NewManager("aaaa", newFakeWG(), nil)
-	dA.Start(&loopback{ep: epA, sw: newSwitch()})
-	dA.Stop()
-	dA.Stop() // must not panic or hang
-}

client/internal/pqkem/message.go

@@ -1,121 +0,0 @@
-package pqkem
-
-import (
-	"crypto/mlkem"
-	"fmt"
-)
-
-// Wire framing for the PQ-KEM exchange. Messages are self-contained, versioned,
-// transport-agnostic byte blobs: the same bytes ride the signalling channel
-// (initial bootstrap) or a data-tunnel packet (rekey). The library only ever sees
-// opaque []byte at the transport seam.
-//
-// Layout (all messages): [type:1][version:1][exchangeID:16][payload...]
-//
-// There is no confirm message: an exchange is acknowledged by the NEXT offer, which
-// carries the acked exchange's id (see OfferMsg.AckID) and — riding the data path
-// under the freshly adopted key — proves that key works.
-
-const (
-	// ProtocolVersion is bumped on any wire-incompatible change; a peer rejects
-	// messages it does not understand rather than misparsing them.
-	ProtocolVersion uint8 = 1
-
-	// ExchangeIDSize identifies one exchange so answers/acks correlate and stale
-	// messages are dropped.
-	ExchangeIDSize = 16
-
-	headerSize = 1 + 1 + ExchangeIDSize
-)
-
-// MsgType tags the two message kinds of the exchange.
-type MsgType uint8
-
-const (
-	MsgOffer MsgType = iota + 1
-	MsgAnswer
-)
-
-// ExchangeID is the per-exchange correlator. The zero value means "none" (an offer
-// that acknowledges nothing, i.e. the first exchange of a connection).
-type ExchangeID [ExchangeIDSize]byte
-
-// OfferMsg carries the initiator's public material (ML-KEM encap key ‖ X25519 pub)
-// and AckID, the id of the previous exchange this offer acknowledges (zero if none).
-type OfferMsg struct {
-	ExchangeID ExchangeID
-	AckID      ExchangeID
-	// KEMOffer is the raw Initiator.Offer() blob (OfferSize bytes).
-	KEMOffer []byte
-}
-
-// AnswerMsg carries the responder's reply (ML-KEM ciphertext ‖ X25519 pub) for the
-// round identified by ExchangeID.
-type AnswerMsg struct {
-	ExchangeID ExchangeID
-	// KEMAnswer is the raw Respond() answer blob (AnswerSize bytes).
-	KEMAnswer []byte
-}
-
-// Encode serialises the offer with its framed header (payload = AckID ‖ KEMOffer).
-func (m *OfferMsg) Encode() ([]byte, error) {
-	if len(m.KEMOffer) != OfferSize {
-		return nil, fmt.Errorf("offer payload: got %d, want %d", len(m.KEMOffer), OfferSize)
-	}
-	payload := make([]byte, 0, ExchangeIDSize+OfferSize)
-	payload = append(payload, m.AckID[:]...)
-	payload = append(payload, m.KEMOffer...)
-	return frame(MsgOffer, m.ExchangeID, payload), nil
-}
-
-// Encode serialises the answer with its framed header.
-func (m *AnswerMsg) Encode() ([]byte, error) {
-	if len(m.KEMAnswer) != AnswerSize {
-		return nil, fmt.Errorf("answer payload: got %d, want %d", len(m.KEMAnswer), AnswerSize)
-	}
-	return frame(MsgAnswer, m.ExchangeID, m.KEMAnswer), nil
-}
-
-// Decode parses a framed message into one of *OfferMsg / *AnswerMsg.
-func Decode(buf []byte) (MsgType, any, error) {
-	if len(buf) < headerSize {
-		return 0, nil, fmt.Errorf("message too short: %d bytes", len(buf))
-	}
-	typ := MsgType(buf[0])
-	if ver := buf[1]; ver != ProtocolVersion {
-		return typ, nil, fmt.Errorf("unsupported protocol version %d (want %d)", ver, ProtocolVersion)
-	}
-
-	var id ExchangeID
-	copy(id[:], buf[2:headerSize])
-	payload := buf[headerSize:]
-
-	switch typ {
-	case MsgOffer:
-		if len(payload) != ExchangeIDSize+OfferSize {
-			return typ, nil, fmt.Errorf("offer payload: got %d, want %d", len(payload), ExchangeIDSize+OfferSize)
-		}
-		var ack ExchangeID
-		copy(ack[:], payload[:ExchangeIDSize])
-		return typ, &OfferMsg{ExchangeID: id, AckID: ack, KEMOffer: payload[ExchangeIDSize:]}, nil
-	case MsgAnswer:
-		if len(payload) != AnswerSize {
-			return typ, nil, fmt.Errorf("answer payload: got %d, want %d", len(payload), AnswerSize)
-		}
-		return typ, &AnswerMsg{ExchangeID: id, KEMAnswer: payload}, nil
-	default:
-		return typ, nil, fmt.Errorf("unknown message type %d", typ)
-	}
-}
-
-func frame(typ MsgType, id ExchangeID, payload []byte) []byte {
-	buf := make([]byte, headerSize+len(payload))
-	buf[0] = byte(typ)
-	buf[1] = ProtocolVersion
-	copy(buf[2:], id[:])
-	copy(buf[headerSize:], payload)
-	return buf
-}
-
-// compile-time assurance the KEM blob sizes referenced here stay in sync with kem.go.
-var _ = [1]struct{}{}[OfferSize-(32+mlkem.EncapsulationKeySize768)]

client/internal/pqkem/message_test.go

@@ -1,57 +0,0 @@
-package pqkem
-
-import (
-	"testing"
-
-	"github.com/stretchr/testify/require"
-)
-
-func TestMessageRoundTrip(t *testing.T) {
-	init, err := NewInitiator()
-	require.NoError(t, err)
-	answer, _, err := Respond(init.Offer(), Binding{LocalID: wgB, RemoteID: wgA})
-	require.NoError(t, err)
-
-	id := ExchangeID{1, 2, 3, 4}
-	ack := ExchangeID{9, 9, 9}
-
-	offBytes, err := (&OfferMsg{ExchangeID: id, AckID: ack, KEMOffer: init.Offer()}).Encode()
-	require.NoError(t, err)
-	typ, decoded, err := Decode(offBytes)
-	require.NoError(t, err)
-	require.Equal(t, MsgOffer, typ)
-	require.Equal(t, id, decoded.(*OfferMsg).ExchangeID)
-	require.Equal(t, ack, decoded.(*OfferMsg).AckID)
-	require.Equal(t, init.Offer(), decoded.(*OfferMsg).KEMOffer)
-
-	ansBytes, err := (&AnswerMsg{ExchangeID: id, KEMAnswer: answer}).Encode()
-	require.NoError(t, err)
-	typ, decoded, err = Decode(ansBytes)
-	require.NoError(t, err)
-	require.Equal(t, MsgAnswer, typ)
-	require.Equal(t, answer, decoded.(*AnswerMsg).KEMAnswer)
-}
-
-func TestDecodeRejects(t *testing.T) {
-	// too short
-	_, _, err := Decode([]byte{1, 1})
-	require.Error(t, err)
-
-	// wrong version
-	bad := make([]byte, headerSize+ExchangeIDSize+OfferSize)
-	bad[0] = byte(MsgOffer)
-	bad[1] = ProtocolVersion + 1
-	_, _, err = Decode(bad)
-	require.Error(t, err)
-
-	// unknown type
-	bad2 := make([]byte, headerSize)
-	bad2[0] = 99
-	bad2[1] = ProtocolVersion
-	_, _, err = Decode(bad2)
-	require.Error(t, err)
-
-	// offer with wrong payload size
-	_, err = (&OfferMsg{KEMOffer: []byte{1, 2, 3}}).Encode()
-	require.Error(t, err)
-}

client/internal/pqkem_adapter.go

@@ -1,160 +0,0 @@
-package internal
-
-import (
-	"net/netip"
-	"time"
-
-	log "github.com/sirupsen/logrus"
-	"golang.zx2c4.com/wireguard/wgctrl/wgtypes"
-
-	"github.com/netbirdio/netbird/client/internal/pqkem"
-)
-
-// pqPresharedKeySetter is the subset of the WireGuard interface the ML-KEM callback
-// needs: programming a peer's preshared key. *iface.WGIface satisfies it.
-type pqPresharedKeySetter interface {
-	SetPresharedKey(peerKey string, psk wgtypes.Key, updateOnly bool) error
-}
-
-// pqCallbackHandler programs the derived PQ PSK onto the WireGuard peer. It is the
-// engine-side implementation of pqkem.CallbackHandler.
-type pqCallbackHandler struct {
-	wg pqPresharedKeySetter
-	// reoffer re-bootstraps the KEM over Signal for a peer (a fresh signalling offer)
-	// to recover from a persistent data-path rekey failure. Nil disables recovery.
-	reoffer func(remoteKey string)
-}
-
-// OnNewPSKReady programs the freshly derived PSK for the peer (updateOnly: a no-op
-// if the peer is not present, mirroring Rosenpass).
-func (h pqCallbackHandler) OnNewPSKReady(remoteID pqkem.RemoteID, psk pqkem.PSK) error {
-	// updateOnly: applies to an already-configured peer (rotation). At bootstrap the
-	// peer is not configured yet, so this is a no-op there and the PSK is instead
-	// pulled at peer-config time (pqHandshaker.PSK / conn.presharedKey).
-	log.Tracef("pqkem: programming PSK for peer %s", remoteID)
-	return h.wg.SetPresharedKey(string(remoteID), wgtypes.Key(psk), true)
-}
-
-// OnRekeyFailed reports a failed PQ (re)key convergence and re-bootstraps the KEM over
-// Signal to recover: a fresh signalling offer starts a new exchange that overwrites the
-// stalled PSK on both sides, resyncing after a persistent data-path desync. The tunnel
-// stays up on the previous PSK meanwhile (the Signal channel is independent of the
-// broken data path).
-func (h pqCallbackHandler) OnRekeyFailed(remoteID pqkem.RemoteID) error {
-	log.Warnf("pqkem: post-quantum rekey failed for peer %s, re-bootstrapping over signal", remoteID)
-	if h.reoffer != nil {
-		h.reoffer(string(remoteID))
-	}
-	return nil
-}
-
-// pqHandshaker adapts the pqkem manager to peer.PQHandshaker (string peer keys),
-// wiring the host's signalling offers/answers to the KEM exchange.
-type pqHandshaker struct {
-	mgr *pqkem.Manager
-}
-
-// announcedPort is the PQ data-path port to advertise to peers. It is omitted (0) when
-// the manager is on DefaultPort, since peers assume the default when no port is sent;
-// only a non-default (collision-forced) port is announced explicitly.
-func (p pqHandshaker) announcedPort() uint16 {
-	if port := p.mgr.LocalPort(); port != DefaultPort {
-		return uint16(port)
-	}
-	return 0
-}
-
-// OfferPayload builds the KEM offer to attach to an outgoing signalling offer for the
-// peer, plus the data-path port to announce (0 when on DefaultPort). Payload is nil when
-// this side has no offer to send.
-func (p pqHandshaker) OfferPayload(remoteKey string) ([]byte, uint16) {
-	payload, err := p.mgr.SignalOffer(pqkem.RemoteID(remoteKey))
-	if err != nil {
-		log.Warnf("pqkem: build offer for %s: %v", remoteKey, err)
-	}
-	return payload, p.announcedPort()
-}
-
-// ShouldSendBootstrapOffer reports whether the controller should reply to the peer's
-// KEM-less offer with its own bootstrap offer instead of an answer.
-func (p pqHandshaker) ShouldSendBootstrapOffer(remoteKey string) bool {
-	return p.mgr.ShouldSendBootstrapOffer(pqkem.RemoteID(remoteKey))
-}
-
-// AnswerPayload processes a received KEM offer (nil when absent) and returns the KEM
-// answer to attach to the outgoing signalling answer, plus the data-path port to announce
-// (0 when on DefaultPort). An empty offer is treated as a capability signal.
-func (p pqHandshaker) AnswerPayload(remoteKey string, recvOffer []byte) ([]byte, uint16) {
-	if len(recvOffer) == 0 {
-		// Capability signal (responder side): the KEM offer flows initiator->responder,
-		// so if we are the responder for this peer (it is the KEM initiator by role) an
-		// empty offer means it does not run the KEM. If we are the initiator, an empty
-		// offer is normal — the peer is the responder and puts its material in the
-		// answer — so we must not flag it.
-		if !p.mgr.IsInitiator(pqkem.RemoteID(remoteKey)) {
-			p.mgr.MarkNonCapable(pqkem.RemoteID(remoteKey))
-		}
-		return nil, p.announcedPort()
-	}
-	payload, err := p.mgr.SignalOnOffer(pqkem.RemoteID(remoteKey), recvOffer)
-	if err != nil {
-		log.Warnf("pqkem: build answer for %s: %v", remoteKey, err)
-	}
-	return payload, p.announcedPort()
-}
-
-// OnAnswer feeds a received KEM answer (nil when absent) into the exchange. An empty
-// answer to our offer is treated as a capability signal on the initiator side.
-func (p pqHandshaker) OnAnswer(remoteKey string, recvAnswer []byte) {
-	if len(recvAnswer) == 0 {
-		// Capability signal (initiator side): the KEM answer flows responder->initiator,
-		// so an empty answer to our offer means the peer does not run the KEM — mark it
-		// non-capable to stop offering (no failure/reoffer storm). Only meaningful when
-		// we are the initiator: as the responder we also receive an (empty) answer to
-		// our own non-KEM offer from a perfectly capable peer, which must not be flagged.
-		if p.mgr.IsInitiator(pqkem.RemoteID(remoteKey)) {
-			p.mgr.MarkNonCapable(pqkem.RemoteID(remoteKey))
-		}
-		return
-	}
-	if err := p.mgr.SignalOnAnswer(pqkem.RemoteID(remoteKey), recvAnswer); err != nil {
-		log.Warnf("pqkem: process answer from %s: %v", remoteKey, err)
-	}
-}
-
-// PSK exposes the peer's derived PSK for the conn to program at WG peer-config time.
-func (p pqHandshaker) PSK(remoteKey string) (wgtypes.Key, bool) {
-	psk, ok := p.mgr.PSK(pqkem.RemoteID(remoteKey))
-	if !ok {
-		return wgtypes.Key{}, false
-	}
-	return wgtypes.Key(psk), true
-}
-
-// SetRemoteAddr registers the peer's data-path endpoint learned from signalling. A
-// zero port means the peer omitted it (it is on DefaultPort), so we resolve it here —
-// DefaultPort lives in this package, not in peer. Sends only ever fire once the tunnel
-// is up (clocked by OnDataPathRekeyed), so registering here is safe even before
-// connection-up.
-func (p pqHandshaker) SetRemoteAddr(remoteKey string, addr netip.AddrPort) {
-	if !addr.Addr().IsValid() {
-		return
-	}
-	port := addr.Port()
-	if port == 0 {
-		port = DefaultPort
-	}
-	p.mgr.AddPeer(pqkem.RemoteID(remoteKey), netip.AddrPortFrom(addr.Addr(), port))
-}
-
-// OnDataPathRekeyed clocks the next chained PSK rotation on a fresh WG handshake.
-// sinceActivity is how long ago the peer last exchanged real user data; the manager
-// skips rotation for idle tunnels.
-func (p pqHandshaker) OnDataPathRekeyed(remoteKey string, sinceActivity time.Duration) {
-	p.mgr.OnDataPathRekeyed(pqkem.RemoteID(remoteKey), sinceActivity)
-}
-
-// OnDataPathDown signals the peer's tunnel went down.
-func (p pqHandshaker) OnDataPathDown(remoteKey string) {
-	p.mgr.OnDataPathDown(pqkem.RemoteID(remoteKey))
-}

client/internal/pqkem_adapter_test.go

@@ -1,37 +0,0 @@
-package internal
-
-import (
-	"testing"
-
-	"github.com/stretchr/testify/require"
-
-	"github.com/netbirdio/netbird/client/internal/pqkem"
-)
-
-type pqNoopHandler struct{}
-
-func (pqNoopHandler) OnNewPSKReady(pqkem.RemoteID, pqkem.PSK) error { return nil }
-func (pqNoopHandler) OnRekeyFailed(pqkem.RemoteID) error            { return nil }
-
-// TestPQAdapter_CapabilityRoleAware locks the role-aware capability signal: the KEM
-// payload only flows initiator-offer -> responder-answer, so an empty message in the
-// other direction comes from a perfectly capable peer and must NOT flag it. Only the
-// message that should carry material (the answer we receive as initiator) marks a peer
-// non-capable when empty.
-func TestPQAdapter_CapabilityRoleAware(t *testing.T) {
-	// localID "zzzz" > "aaaa" => this manager is the KEM initiator for peer "aaaa".
-	mgr := pqkem.NewManager("zzzz", pqNoopHandler{}, nil)
-	defer mgr.Stop()
-	h := pqHandshaker{mgr: mgr}
-
-	// An empty OFFER from our peer is normal here: as the initiator's responder it puts
-	// its material in the answer, not the offer. It must not disable our offering.
-	h.AnswerPayload("aaaa", nil)
-	payload, _ := h.OfferPayload("aaaa")
-	require.NotNil(t, payload, "an empty offer from a responder-role peer must not mark it non-capable")
-
-	// An empty ANSWER to our offer means the peer does not run the KEM -> stop offering.
-	h.OnAnswer("aaaa", nil)
-	payload2, _ := h.OfferPayload("aaaa")
-	require.Nil(t, payload2, "an empty answer to our offer marks the peer non-capable, so we stop offering")
-}

client/internal/pqkem_transport.go

@@ -1,76 +0,0 @@
-package internal
-
-import (
-	"fmt"
-	"net"
-	"net/netip"
-
-	log "github.com/sirupsen/logrus"
-)
-
-// DefaultPort is the preferred UDP port for the ML-KEM data-path service, bound on
-// the WG overlay IP. Since each client owns a distinct overlay IP, this port is
-// almost always free, so it need not be announced (peers assume it). A peer only
-// announces Body.mlkemPort when a collision forced it onto a different port.
-const DefaultPort = 51833
-
-// pqTransport is the ML-KEM data-path transport: a dumb UDP socket bound on the WG
-// overlay IP. It implements pqkem.Transport — the manager owns the remoteID<->endpoint
-// routing and drives this socket's lifecycle (Run / Close).
-type pqTransport struct {
-	conn *net.UDPConn
-	port int
-}
-
-// newPQTransport binds a UDP socket on the WG overlay IP, preferring DefaultPort and
-// falling back to an OS-assigned ephemeral port if it is in use. Call it after the WG
-// interface is up so the overlay IP is assigned; when the bound port is not
-// DefaultPort it must be announced to peers via Body.mlkemPort.
-func newPQTransport(overlayIP netip.Addr) (*pqTransport, error) {
-	if !overlayIP.IsValid() {
-		return nil, fmt.Errorf("invalid overlay IP for pqkem transport")
-	}
-	// The WG overlay always carries an IPv4 address (v6 is additive, never standalone),
-	// so the transport binds over IPv4. Unmap first so AsSlice() yields 4 bytes for an
-	// IPv4-mapped IPv6 address (a hardcoded "udp4" would otherwise fail on its 16 bytes).
-	overlayIP = overlayIP.Unmap()
-	ip := net.IP(overlayIP.AsSlice())
-	conn, err := net.ListenUDP("udp4", &net.UDPAddr{IP: ip, Port: DefaultPort})
-	if err != nil {
-		log.Debugf("pqkem: default port %d unavailable on %s (%v), using an ephemeral port", DefaultPort, overlayIP, err)
-		conn, err = net.ListenUDP("udp4", &net.UDPAddr{IP: ip, Port: 0})
-		if err != nil {
-			return nil, fmt.Errorf("bind pqkem udp on overlay %s: %w", overlayIP, err)
-		}
-	}
-	return &pqTransport{conn: conn, port: conn.LocalAddr().(*net.UDPAddr).Port}, nil
-}
-
-// Send implements pqkem.Transport.
-func (t *pqTransport) Send(endpoint netip.AddrPort, msg []byte) error {
-	_, err := t.conn.WriteToUDPAddrPort(msg, endpoint)
-	return err
-}
-
-// LocalPort implements pqkem.Transport.
-func (t *pqTransport) LocalPort() int { return t.port }
-
-// Run implements pqkem.Transport: the receive loop, delivering each datagram as
-// (source endpoint, msg). Exits when the socket is closed.
-func (t *pqTransport) Run(onInbound func(src netip.AddrPort, msg []byte)) {
-	go func() {
-		buf := make([]byte, 2048)
-		for {
-			n, src, err := t.conn.ReadFromUDPAddrPort(buf)
-			if err != nil {
-				return
-			}
-			msg := make([]byte, n)
-			copy(msg, buf[:n])
-			onInbound(src, msg)
-		}
-	}()
-}
-
-// Close implements pqkem.Transport.
-func (t *pqTransport) Close() error { return t.conn.Close() }

client/internal/routemanager/ipfwdstate/ipfwdstate.go

@@ -2,54 +2,183 @@ package ipfwdstate
 
 import (
 	"fmt"
+	"sync"
 
 	log "github.com/sirupsen/logrus"
 
 	"github.com/netbirdio/netbird/client/internal/routemanager/systemops"
 )
 
-// IPForwardingState is a struct that keeps track of the IP forwarding state.
-// todo: read initial state of the IP forwarding from the system and reset the state based on it.
-// todo: separate v4/v6 forwarding state, since the sysctls are independent
-// (net.ipv4.ip_forward vs net.ipv6.conf.all.forwarding). Currently the nftables
-// manager shares one instance between both routers, which works only because
-// EnableIPForwarding enables both sysctls in a single call.
+// IPForwardingState tracks v4 and v6 IP-forwarding sysctl enables with
+// independent refcounts so a v4-only routing setup doesn't flip v6 sysctls.
 type IPForwardingState struct {
-	enabledCounter int
+	mu sync.Mutex
+
+	v4Count int
+	v6Count int
+
+	// routingV4/routingV6 track whether the routing path currently holds a
+	// reference, so repeated EnableRouting calls (one per network-map update)
+	// hold at most one reference per family and an unpaired DisableRouting
+	// can't release references held by DNAT rules.
+	routingV4 bool
+	routingV6 bool
+
+	wgIfaceName string
+	v6Saved     map[string]int
 }
 
-func NewIPForwardingState() *IPForwardingState {
-	return &IPForwardingState{}
+// NewIPForwardingState returns a state tracker for the IP-forwarding sysctls.
+// wgIfaceName is excluded from the per-interface accept_ra handling.
+func NewIPForwardingState(wgIfaceName string) *IPForwardingState {
+	return &IPForwardingState{wgIfaceName: wgIfaceName}
 }
 
-func (f *IPForwardingState) RequestForwarding() error {
-	if f.enabledCounter != 0 {
-		f.enabledCounter++
+// Counts returns the current v4 and v6 refcounts. Intended for diagnostics
+// and tests.
+func (f *IPForwardingState) Counts() (v4, v6 int) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.v4Count, f.v6Count
+}
+
+// RequestRouting takes the forwarding references for the routing path. It is
+// idempotent: while routing already holds a reference, further calls don't
+// increment the refcounts, and a v4-only request releases a previously held v6
+// reference. A v6 sysctl failure is logged and not returned so it can't take
+// down v4 routing (the sysctl may be unwritable, e.g. read-only /proc/sys or
+// IPv6 disabled on the kernel command line); v6 is retried on the next call.
+func (f *IPForwardingState) RequestRouting(v6 bool) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	if !f.routingV4 {
+		if err := f.requestV4(); err != nil {
+			return err
+		}
+		f.routingV4 = true
+	}
+
+	if !v6 {
+		if !f.routingV6 {
+			return nil
+		}
+		f.routingV6 = false
+		return f.releaseV6()
+	}
+
+	if f.routingV6 {
 		return nil
 	}
-
-	if err := systemops.EnableIPForwarding(); err != nil {
-		return fmt.Errorf("failed to enable IP forwarding with sysctl: %w", err)
+	if err := f.requestV6(); err != nil {
+		log.Warnf("enable IPv6 forwarding for routing: %v", err)
+		return nil
 	}
-	f.enabledCounter = 1
-	log.Info("IP forwarding enabled")
-
+	f.routingV6 = true
 	return nil
 }
 
-func (f *IPForwardingState) ReleaseForwarding() error {
-	if f.enabledCounter == 0 {
-		return nil
+// ReleaseRouting releases the references RequestRouting holds. Calls without a
+// held reference are no-ops.
+func (f *IPForwardingState) ReleaseRouting() error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	if f.routingV4 {
+		f.routingV4 = false
+		f.releaseV4()
 	}
-
-	if f.enabledCounter > 1 {
-		f.enabledCounter--
-		return nil
+	if f.routingV6 {
+		f.routingV6 = false
+		return f.releaseV6()
 	}
-
-	// if failed to disable IP forwarding we anyway decrement the counter
-	f.enabledCounter = 0
-
-	// todo call systemops.DisableIPForwarding()
+	return nil
+}
+
+// RequestForwarding enables the family's forwarding sysctl on first request.
+func (f *IPForwardingState) RequestForwarding(v6 bool) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	if v6 {
+		return f.requestV6()
+	}
+	return f.requestV4()
+}
+
+// ReleaseForwarding decrements the family counter. The last v6 release restores
+// what enable captured. v4 stays on: net.ipv4.ip_forward is co-owned by other
+// tooling (docker, k8s, libvirt).
+func (f *IPForwardingState) ReleaseForwarding(v6 bool) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	if v6 {
+		return f.releaseV6()
+	}
+	f.releaseV4()
+	return nil
+}
+
+func (f *IPForwardingState) requestV4() error {
+	if f.v4Count == 0 {
+		if err := systemops.EnableV4IPForwarding(); err != nil {
+			return fmt.Errorf("enable IPv4 forwarding: %w", err)
+		}
+		log.Info("IPv4 forwarding enabled")
+	}
+	f.v4Count++
+	return nil
+}
+
+func (f *IPForwardingState) releaseV4() {
+	if f.v4Count > 0 {
+		f.v4Count--
+	}
+}
+
+func (f *IPForwardingState) requestV6() error {
+	if f.v6Count == 0 {
+		saved, err := systemops.EnableV6IPForwarding(f.wgIfaceName)
+		if err != nil {
+			if rerr := systemops.DisableV6IPForwarding(saved); rerr != nil {
+				log.Warnf("rollback partial v6 sysctls: %v", rerr)
+			}
+			return fmt.Errorf("enable IPv6 forwarding: %w", err)
+		}
+		// A failed restore on a previous release keeps its saved values; those
+		// are the true originals, so keep them over what this enable captured.
+		if f.v6Saved == nil {
+			f.v6Saved = saved
+		} else {
+			for k, v := range saved {
+				if _, ok := f.v6Saved[k]; !ok {
+					f.v6Saved[k] = v
+				}
+			}
+		}
+		log.Info("IPv6 forwarding enabled")
+	}
+	f.v6Count++
+	return nil
+}
+
+func (f *IPForwardingState) releaseV6() error {
+	if f.v6Count == 0 {
+		return nil
+	}
+	f.v6Count--
+	if f.v6Count > 0 {
+		return nil
+	}
+
+	// Keep the saved values on failure so a later release or enable/release
+	// cycle can still restore them; re-restoring an already-restored key is a
+	// no-op since the sysctl already holds the desired value.
+	if err := systemops.DisableV6IPForwarding(f.v6Saved); err != nil {
+		return fmt.Errorf("disable IPv6 forwarding: %w", err)
+	}
+	f.v6Saved = nil
+	log.Info("IPv6 forwarding disabled")
 	return nil
 }

client/internal/routemanager/ipfwdstate/ipfwdstate_privileged_linux_test.go

@@ -0,0 +1,39 @@
+//go:build privileged
+
+package ipfwdstate
+
+import (
+	"testing"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+)
+
+// TestRequestRoutingV6ToV4Transition verifies that a v4-only routing request
+// releases a previously held routing-owned v6 reference without touching
+// references held by DNAT rules.
+func TestRequestRoutingV6ToV4Transition(t *testing.T) {
+	f := NewIPForwardingState("wt-fwd-test")
+
+	require.NoError(t, f.RequestRouting(true), "request routing with v6")
+	v4, v6 := f.Counts()
+	assert.Equal(t, 1, v4, "v4 reference held")
+	assert.Equal(t, 1, v6, "v6 reference held")
+
+	require.NoError(t, f.RequestRouting(false), "request routing v4-only")
+	v4, v6 = f.Counts()
+	assert.Equal(t, 1, v4, "v4 reference kept")
+	assert.Equal(t, 0, v6, "routing-owned v6 reference released")
+
+	// A DNAT-held reference survives a v4-only routing request.
+	require.NoError(t, f.RequestForwarding(true), "dnat v6 reference")
+	require.NoError(t, f.RequestRouting(false), "repeat v4-only request")
+	_, v6 = f.Counts()
+	assert.Equal(t, 1, v6, "dnat-held v6 reference survives")
+	require.NoError(t, f.ReleaseForwarding(true), "release dnat v6 reference")
+
+	require.NoError(t, f.ReleaseRouting(), "release routing")
+	v4, v6 = f.Counts()
+	assert.Equal(t, 0, v4, "all v4 references released")
+	assert.Equal(t, 0, v6, "all v6 references released")
+}

client/internal/routemanager/sysctl/sysctl_linux.go

@@ -58,11 +58,7 @@ func Setup(wgIface iface) (map[string]int, error) {
 			continue
 		}
 
-		// Escape '%' and '.' so they survive the dot-to-slash conversion in Set()
-		safeName := strings.ReplaceAll(intf.Name, "%", percentEscape)
-		safeName = strings.ReplaceAll(safeName, ".", dotEscape)
-
-		i := fmt.Sprintf(rpFilterInterfacePath, safeName)
+		i := fmt.Sprintf(rpFilterInterfacePath, EscapeInterfaceName(intf.Name))
 		oldVal, err := Set(i, 2, true)
 		if err != nil {
 			result = multierror.Append(result, err)
@@ -74,6 +70,13 @@ func Setup(wgIface iface) (map[string]int, error) {
 	return keys, nberrors.FormatErrorOrNil(result)
 }
 
+// EscapeInterfaceName escapes '%' and '.' in an interface name (e.g. VLANs
+// like eth0.100) so the name survives the dot-to-slash conversion in Set.
+func EscapeInterfaceName(name string) string {
+	safe := strings.ReplaceAll(name, "%", percentEscape)
+	return strings.ReplaceAll(safe, ".", dotEscape)
+}
+
 // Set sets a sysctl configuration, if onlyIfOne is true it will only set the new value if it's set to 1
 func Set(key string, desiredValue int, onlyIfOne bool) (int, error) {
 	path := strings.ReplaceAll(key, ".", "/")

client/internal/routemanager/systemops/systemops_android.go

@@ -32,8 +32,17 @@ func (r *SysOps) removeFromRouteTable(netip.Prefix, Nexthop) error {
 	return nil
 }
 
-func EnableIPForwarding() error {
-	log.Infof("Enable IP forwarding is not implemented on %s", runtime.GOOS)
+func EnableV4IPForwarding() error {
+	log.Infof("Enable IPv4 forwarding is not implemented on %s", runtime.GOOS)
+	return nil
+}
+
+func EnableV6IPForwarding(string) (map[string]int, error) {
+	log.Infof("Enable IPv6 forwarding is not implemented on %s", runtime.GOOS)
+	return map[string]int{}, nil
+}
+
+func DisableV6IPForwarding(map[string]int) error {
 	return nil
 }
 

client/internal/routemanager/systemops/systemops_ios.go

@@ -58,8 +58,17 @@ func (r *SysOps) removeFromRouteTable(netip.Prefix, Nexthop) error {
 	return nil
 }
 
-func EnableIPForwarding() error {
-	log.Infof("Enable IP forwarding is not implemented on %s", runtime.GOOS)
+func EnableV4IPForwarding() error {
+	log.Infof("Enable IPv4 forwarding is not implemented on %s", runtime.GOOS)
+	return nil
+}
+
+func EnableV6IPForwarding(string) (map[string]int, error) {
+	log.Infof("Enable IPv6 forwarding is not implemented on %s", runtime.GOOS)
+	return map[string]int{}, nil
+}
+
+func DisableV6IPForwarding(map[string]int) error {
 	return nil
 }
 

client/internal/routemanager/systemops/systemops_linux.go

@@ -763,13 +763,10 @@ func flushRoutes(tableID, family int) error {
 	return nberrors.FormatErrorOrNil(result)
 }
 
-func EnableIPForwarding() error {
+func EnableV4IPForwarding() error {
 	if _, err := sysctl.Set(ipv4ForwardingPath, 1, false); err != nil {
 		return err
 	}
-	if _, err := sysctl.Set(ipv6ForwardingPath, 1, false); err != nil {
-		log.Warnf("failed to enable IPv6 forwarding: %v", err)
-	}
 	return nil
 }
 

client/internal/routemanager/systemops/systemops_nonlinux.go

@@ -43,8 +43,17 @@ func (r *SysOps) RemoveVPNRoute(prefix netip.Prefix, intf *net.Interface) error
 	return r.genericRemoveVPNRoute(prefix, intf)
 }
 
-func EnableIPForwarding() error {
-	log.Infof("Enable IP forwarding is not implemented on %s", runtime.GOOS)
+func EnableV4IPForwarding() error {
+	log.Infof("Enable IPv4 forwarding is not implemented on %s", runtime.GOOS)
+	return nil
+}
+
+func EnableV6IPForwarding(string) (map[string]int, error) {
+	log.Infof("Enable IPv6 forwarding is not implemented on %s", runtime.GOOS)
+	return map[string]int{}, nil
+}
+
+func DisableV6IPForwarding(map[string]int) error {
 	return nil
 }
 

client/internal/routemanager/systemops/v6forwarding_linux.go

@@ -0,0 +1,92 @@
+//go:build !android
+
+package systemops
+
+import (
+	"fmt"
+	"net"
+	"os"
+
+	"github.com/hashicorp/go-multierror"
+	log "github.com/sirupsen/logrus"
+
+	nberrors "github.com/netbirdio/netbird/client/errors"
+	"github.com/netbirdio/netbird/client/internal/routemanager/sysctl"
+)
+
+const (
+	// 1 (default) accepts RAs only while forwarding is off; 2 keeps RA
+	// acceptance on regardless, so RA-installed host defaults survive our
+	// v6 forwarding flip.
+	acceptRAInterfacePath  = "net.ipv6.conf.%s.accept_ra"
+	acceptRADefaultPath    = "net.ipv6.conf.default.accept_ra"
+	acceptRAProcPathFormat = "/proc/sys/net/ipv6/conf/%s/accept_ra"
+)
+
+// EnableV6IPForwarding bumps accept_ra=2 on host v6 interfaces before flipping
+// forwarding=1, so RA-installed host defaults survive. Returns the prior values
+// of sysctls we actually changed; entries already at the target are omitted.
+func EnableV6IPForwarding(wgIfaceName string) (map[string]int, error) {
+	saved := map[string]int{}
+	bumpAcceptRA(saved, wgIfaceName)
+
+	oldVal, err := sysctl.Set(ipv6ForwardingPath, 1, false)
+	if err != nil {
+		return saved, err
+	}
+	if oldVal != 1 {
+		saved[ipv6ForwardingPath] = oldVal
+	}
+	return saved, nil
+}
+
+// DisableV6IPForwarding restores what EnableV6IPForwarding captured.
+func DisableV6IPForwarding(saved map[string]int) error {
+	var result *multierror.Error
+	for key, value := range saved {
+		if _, err := sysctl.Set(key, value, false); err != nil {
+			result = multierror.Append(result, fmt.Errorf("restore %s: %w", key, err))
+		}
+	}
+	return nberrors.FormatErrorOrNil(result)
+}
+
+func bumpAcceptRA(saved map[string]int, wgIfaceName string) {
+	// Also bump conf.default so interfaces created while forwarding is on
+	// (hotplug, new Wi-Fi/dock) inherit accept_ra=2 and keep accepting RAs.
+	bumpAcceptRAKey(saved, acceptRADefaultPath)
+
+	interfaces, err := net.Interfaces()
+	if err != nil {
+		log.Warnf("list interfaces for accept_ra: %v", err)
+		return
+	}
+	for _, intf := range interfaces {
+		if intf.Name == "lo" || intf.Name == wgIfaceName {
+			continue
+		}
+		bumpAcceptRAForInterface(saved, intf.Name)
+	}
+}
+
+func bumpAcceptRAForInterface(saved map[string]int, name string) {
+	// Build procfs path from name, not the dotted key: VLAN names like eth0.100.
+	if _, err := os.Stat(fmt.Sprintf(acceptRAProcPathFormat, name)); err != nil {
+		return
+	}
+	bumpAcceptRAKey(saved, fmt.Sprintf(acceptRAInterfacePath, sysctl.EscapeInterfaceName(name)))
+}
+
+func bumpAcceptRAKey(saved map[string]int, key string) {
+	// onlyIfOne=true: leave admin overrides (0, 2) alone.
+	oldVal, err := sysctl.Set(key, 2, true)
+	if err != nil {
+		log.Warnf("bump %s: %v", key, err)
+		return
+	}
+	// With onlyIfOne, a write only happened when the old value was 1; values
+	// left untouched (0, 2) must not be recorded for restore.
+	if oldVal == 1 {
+		saved[key] = oldVal
+	}
+}

client/ios/NetBirdSDK/client.go

@@ -14,6 +14,7 @@ import (
 
 	log "github.com/sirupsen/logrus"
 
+	nbAnonymize "github.com/netbirdio/netbird/client/anonymize"
 	"github.com/netbirdio/netbird/client/internal"
 	"github.com/netbirdio/netbird/client/internal/auth"
 	"github.com/netbirdio/netbird/client/internal/debug"
@@ -28,6 +29,13 @@ import (
 	types "github.com/netbirdio/netbird/upload-server/types"
 )
 
+// AnonymizeLevelDefault and AnonymizeLevelStrict are the accepted
+// anonymizeLevel values for DebugBundle.
+const (
+	AnonymizeLevelDefault = nbAnonymize.LevelDefaultString
+	AnonymizeLevelStrict  = nbAnonymize.LevelStrictString
+)
+
 // ConnectionListener export internal Listener for mobile
 type ConnectionListener interface {
 	peer.Listener
@@ -200,8 +208,10 @@ func (c *Client) Stop() {
 // DebugBundle generates a debug bundle, uploads it and returns the upload key.
 // It works with or without a running engine: when the engine is up it reuses
 // the live config, sync response and client metrics; otherwise it loads the
-// config from disk (or the preloaded tvOS config).
-func (c *Client) DebugBundle(anonymize bool) (string, error) {
+// config from disk (or the preloaded tvOS config). anonymizeLevel is "default"
+// or "strict"; strict also anonymizes internal IP ranges, peer names, and
+// WireGuard public keys, and implies anonymize.
+func (c *Client) DebugBundle(anonymize bool, anonymizeLevel string) (string, error) {
 	cfg, cc := c.stateSnapshot()
 
 	// If the engine hasn't been started, load config so we can reach management.
@@ -251,6 +261,7 @@ func (c *Client) DebugBundle(anonymize bool) (string, error) {
 		deps,
 		debug.BundleConfig{
 			Anonymize:         anonymize,
+			AnonymizeLevel:    nbAnonymize.ParseLevel(anonymizeLevel),
 			IncludeSystemInfo: true,
 		},
 	)

client/proto/daemon.pb.go

@@ -2781,6 +2781,11 @@ type DebugBundleRequest struct {
 	// untrusted TLS certificate. Restricted to privileged callers; for
 	// self-hosted upload servers.
 	UploadInsecure bool `protobuf:"varint,7,opt,name=uploadInsecure,proto3" json:"uploadInsecure,omitempty"`
+	// anonymizeLevel selects how much the anonymizer redacts: "default"
+	// (or empty) keeps internal IP ranges, "strict" also anonymizes them.
+	// Unknown values are treated as "strict". Only meaningful with anonymize;
+	// "strict" implies it.
+	AnonymizeLevel string `protobuf:"bytes,8,opt,name=anonymizeLevel,proto3" json:"anonymizeLevel,omitempty"`
 	unknownFields  protoimpl.UnknownFields
 	sizeCache      protoimpl.SizeCache
 }
@@ -2857,6 +2862,13 @@ func (x *DebugBundleRequest) GetUploadInsecure() bool {
 	return false
 }
 
+func (x *DebugBundleRequest) GetAnonymizeLevel() string {
+	if x != nil {
+		return x.AnonymizeLevel
+	}
+	return ""
+}
+
 type DebugBundleResponse struct {
 	state               protoimpl.MessageState `protogen:"open.v1"`
 	Path                string                 `protobuf:"bytes,1,opt,name=path,proto3" json:"path,omitempty"`
@@ -5616,9 +5628,13 @@ func (x *GetPeerSSHHostKeyResponse) GetFound() bool {
 type RequestJWTAuthRequest struct {
 	state protoimpl.MessageState `protogen:"open.v1"`
 	// hint for OIDC login_hint parameter (typically email address)
-	Hint          *string `protobuf:"bytes,1,opt,name=hint,proto3,oneof" json:"hint,omitempty"`
-	unknownFields protoimpl.UnknownFields
-	sizeCache     protoimpl.SizeCache
+	Hint *string `protobuf:"bytes,1,opt,name=hint,proto3,oneof" json:"hint,omitempty"`
+	// hasGraphicalSession tells the daemon that the caller has a graphical session,
+	// which decides whether PKCE or the device code flow is preferred. The daemon
+	// cannot detect this itself: it does not inherit the session environment.
+	HasGraphicalSession bool `protobuf:"varint,2,opt,name=hasGraphicalSession,proto3" json:"hasGraphicalSession,omitempty"`
+	unknownFields       protoimpl.UnknownFields
+	sizeCache           protoimpl.SizeCache
 }
 
 func (x *RequestJWTAuthRequest) Reset() {
@@ -5658,6 +5674,13 @@ func (x *RequestJWTAuthRequest) GetHint() string {
 	return ""
 }
 
+func (x *RequestJWTAuthRequest) GetHasGraphicalSession() bool {
+	if x != nil {
+		return x.HasGraphicalSession
+	}
+	return false
+}
+
 // RequestJWTAuthResponse contains authentication flow information
 type RequestJWTAuthResponse struct {
 	state protoimpl.MessageState `protogen:"open.v1"`
@@ -5882,9 +5905,13 @@ type RequestExtendAuthSessionRequest struct {
 	state protoimpl.MessageState `protogen:"open.v1"`
 	// Optional OIDC login_hint (typically the user's email) to pre-fill the
 	// IdP login form.
-	Hint          *string `protobuf:"bytes,1,opt,name=hint,proto3,oneof" json:"hint,omitempty"`
-	unknownFields protoimpl.UnknownFields
-	sizeCache     protoimpl.SizeCache
+	Hint *string `protobuf:"bytes,1,opt,name=hint,proto3,oneof" json:"hint,omitempty"`
+	// hasGraphicalSession tells the daemon that the caller has a graphical session,
+	// which decides whether PKCE or the device code flow is preferred. The daemon
+	// cannot detect this itself: it does not inherit the session environment.
+	HasGraphicalSession bool `protobuf:"varint,2,opt,name=hasGraphicalSession,proto3" json:"hasGraphicalSession,omitempty"`
+	unknownFields       protoimpl.UnknownFields
+	sizeCache           protoimpl.SizeCache
 }
 
 func (x *RequestExtendAuthSessionRequest) Reset() {
@@ -5924,6 +5951,13 @@ func (x *RequestExtendAuthSessionRequest) GetHint() string {
 	return ""
 }
 
+func (x *RequestExtendAuthSessionRequest) GetHasGraphicalSession() bool {
+	if x != nil {
+		return x.HasGraphicalSession
+	}
+	return false
+}
+
 // RequestExtendAuthSessionResponse carries the verification URI the UI
 // should open in a browser. The daemon retains the flow state and resolves
 // it via WaitExtendAuthSession.
@@ -7253,7 +7287,7 @@ const file_daemon_proto_rawDesc = "" +
 	"\x12translatedHostname\x18\x04 \x01(\tR\x12translatedHostname\x128\n" +
 	"\x0etranslatedPort\x18\x05 \x01(\v2\x10.daemon.PortInfoR\x0etranslatedPort\"G\n" +
 	"\x17ForwardingRulesResponse\x12,\n" +
-	"\x05rules\x18\x01 \x03(\v2\x16.daemon.ForwardingRuleR\x05rules\"\xdc\x01\n" +
+	"\x05rules\x18\x01 \x03(\v2\x16.daemon.ForwardingRuleR\x05rules\"\x84\x02\n" +
 	"\x12DebugBundleRequest\x12\x1c\n" +
 	"\tanonymize\x18\x01 \x01(\bR\tanonymize\x12\x1e\n" +
 	"\n" +
@@ -7264,7 +7298,8 @@ const file_daemon_proto_rawDesc = "" +
 	"\n" +
 	"cliVersion\x18\x06 \x01(\tR\n" +
 	"cliVersion\x12&\n" +
-	"\x0euploadInsecure\x18\a \x01(\bR\x0euploadInsecure\"}\n" +
+	"\x0euploadInsecure\x18\a \x01(\bR\x0euploadInsecure\x12&\n" +
+	"\x0eanonymizeLevel\x18\b \x01(\tR\x0eanonymizeLevel\"}\n" +
 	"\x13DebugBundleResponse\x12\x12\n" +
 	"\x04path\x18\x01 \x01(\tR\x04path\x12 \n" +
 	"\vuploadedKey\x18\x02 \x01(\tR\vuploadedKey\x120\n" +
@@ -7490,9 +7525,10 @@ const file_daemon_proto_rawDesc = "" +
 	"sshHostKey\x12\x16\n" +
 	"\x06peerIP\x18\x02 \x01(\tR\x06peerIP\x12\x1a\n" +
 	"\bpeerFQDN\x18\x03 \x01(\tR\bpeerFQDN\x12\x14\n" +
-	"\x05found\x18\x04 \x01(\bR\x05found\"9\n" +
+	"\x05found\x18\x04 \x01(\bR\x05found\"k\n" +
 	"\x15RequestJWTAuthRequest\x12\x17\n" +
-	"\x04hint\x18\x01 \x01(\tH\x00R\x04hint\x88\x01\x01B\a\n" +
+	"\x04hint\x18\x01 \x01(\tH\x00R\x04hint\x88\x01\x01\x120\n" +
+	"\x13hasGraphicalSession\x18\x02 \x01(\bR\x13hasGraphicalSessionB\a\n" +
 	"\x05_hint\"\x9a\x02\n" +
 	"\x16RequestJWTAuthResponse\x12(\n" +
 	"\x0fverificationURI\x18\x01 \x01(\tR\x0fverificationURI\x128\n" +
@@ -7512,9 +7548,10 @@ const file_daemon_proto_rawDesc = "" +
 	"\x14WaitJWTTokenResponse\x12\x14\n" +
 	"\x05token\x18\x01 \x01(\tR\x05token\x12\x1c\n" +
 	"\ttokenType\x18\x02 \x01(\tR\ttokenType\x12\x1c\n" +
-	"\texpiresIn\x18\x03 \x01(\x03R\texpiresIn\"C\n" +
+	"\texpiresIn\x18\x03 \x01(\x03R\texpiresIn\"u\n" +
 	"\x1fRequestExtendAuthSessionRequest\x12\x17\n" +
-	"\x04hint\x18\x01 \x01(\tH\x00R\x04hint\x88\x01\x01B\a\n" +
+	"\x04hint\x18\x01 \x01(\tH\x00R\x04hint\x88\x01\x01\x120\n" +
+	"\x13hasGraphicalSession\x18\x02 \x01(\bR\x13hasGraphicalSessionB\a\n" +
 	"\x05_hint\"\xe0\x01\n" +
 	" RequestExtendAuthSessionResponse\x12(\n" +
 	"\x0fverificationURI\x18\x01 \x01(\tR\x0fverificationURI\x128\n" +

client/proto/daemon.proto

@@ -540,6 +540,11 @@ message DebugBundleRequest {
   // untrusted TLS certificate. Restricted to privileged callers; for
   // self-hosted upload servers.
   bool uploadInsecure = 7;
+  // anonymizeLevel selects how much the anonymizer redacts: "default"
+  // (or empty) keeps internal IP ranges, "strict" also anonymizes them.
+  // Unknown values are treated as "strict". Only meaningful with anonymize;
+  // "strict" implies it.
+  string anonymizeLevel = 8;
 }
 
 message DebugBundleResponse {
@@ -889,6 +894,10 @@ message GetPeerSSHHostKeyResponse {
 message RequestJWTAuthRequest {
   // hint for OIDC login_hint parameter (typically email address)
   optional string hint = 1;
+  // hasGraphicalSession tells the daemon that the caller has a graphical session,
+  // which decides whether PKCE or the device code flow is preferred. The daemon
+  // cannot detect this itself: it does not inherit the session environment.
+  bool hasGraphicalSession = 2;
 }
 
 // RequestJWTAuthResponse contains authentication flow information
@@ -932,6 +941,10 @@ message RequestExtendAuthSessionRequest {
   // Optional OIDC login_hint (typically the user's email) to pre-fill the
   // IdP login form.
   optional string hint = 1;
+  // hasGraphicalSession tells the daemon that the caller has a graphical session,
+  // which decides whether PKCE or the device code flow is preferred. The daemon
+  // cannot detect this itself: it does not inherit the session environment.
+  bool hasGraphicalSession = 2;
 }
 
 // RequestExtendAuthSessionResponse carries the verification URI the UI

client/server/debug.go

@@ -16,6 +16,7 @@ import (
 	"google.golang.org/grpc/codes"
 	gstatus "google.golang.org/grpc/status"
 
+	"github.com/netbirdio/netbird/client/anonymize"
 	"github.com/netbirdio/netbird/client/internal/debug"
 	"github.com/netbirdio/netbird/client/internal/ipcauth"
 	"github.com/netbirdio/netbird/client/proto"
@@ -122,6 +123,7 @@ func (s *Server) generateDebugBundle(req *proto.DebugBundleRequest, uiOpener deb
 		},
 		debug.BundleConfig{
 			Anonymize:         req.GetAnonymize(),
+			AnonymizeLevel:    anonymize.ParseLevel(req.GetAnonymizeLevel()),
 			IncludeSystemInfo: req.GetSystemInfo(),
 			LogFileCount:      req.GetLogFileCount(),
 		},

client/server/server.go

@@ -682,6 +682,11 @@ func (s *Server) Login(callerCtx context.Context, msg *proto.LoginRequest) (*pro
 		oAuthFlow, err := auth.NewOAuthFlow(ctx, config, msg.IsUnixDesktopClient, false, hint)
 		if err != nil {
 			state.Set(internal.StatusLoginFailed)
+			// enrolling a device is the one flow a setup key can replace. NotFound so the CLI
+			// stops its backoff loop and shows this instead of retrying a permanent condition.
+			if auth.IsSSOUnavailable(err) {
+				return nil, gstatus.Error(codes.NotFound, auth.WithSetupKeyAdvice(err).Error())
+			}
 			return nil, err
 		}
 
@@ -1723,8 +1728,8 @@ func (s *Server) RequestJWTAuth(
 		hint = profilemanager.GetLoginHint()
 	}
 
-	isDesktop := isUnixRunningDesktop()
-	oAuthFlow, err := auth.NewOAuthFlow(ctx, config, isDesktop, false, hint)
+	// the daemon has no graphical session of its own, only the caller can answer this
+	oAuthFlow, err := auth.NewOAuthFlow(ctx, config, msg.GetHasGraphicalSession(), false, hint)
 	if err != nil {
 		return nil, gstatus.Errorf(codes.Internal, "failed to create OAuth flow: %v", err)
 	}
@@ -1827,8 +1832,8 @@ func (s *Server) RequestExtendAuthSession(
 		hint = profilemanager.GetLoginHint()
 	}
 
-	isDesktop := isUnixRunningDesktop()
-	oAuthFlow, err := auth.NewOAuthFlow(ctx, config, isDesktop, false, hint)
+	// the daemon has no graphical session of its own, only the caller can answer this
+	oAuthFlow, err := auth.NewOAuthFlow(ctx, config, msg.GetHasGraphicalSession(), false, hint)
 	if err != nil {
 		return nil, gstatus.Errorf(codes.Internal, "failed to create OAuth flow: %v", err)
 	}
@@ -2000,13 +2005,6 @@ func (s *Server) ExposeService(req *proto.ExposeServiceRequest, srv proto.Daemon
 	return nil
 }
 
-func isUnixRunningDesktop() bool {
-	if runtime.GOOS != "linux" && runtime.GOOS != "freebsd" {
-		return false
-	}
-	return os.Getenv("DESKTOP_SESSION") != "" || os.Getenv("XDG_CURRENT_DESKTOP") != ""
-}
-
 func (s *Server) runProbes(ctx context.Context, waitForProbeResult bool) {
 	if s.connectClient == nil {
 		return

client/ssh/common.go

@@ -13,6 +13,7 @@ import (
 	"golang.org/x/crypto/ssh"
 
 	"github.com/netbirdio/netbird/client/proto"
+	"github.com/netbirdio/netbird/util"
 )
 
 const (
@@ -92,7 +93,8 @@ func printAuthInstructions(stderr io.Writer, authResponse *proto.RequestJWTAuthR
 
 // RequestJWTToken requests or retrieves a JWT token for SSH authentication
 func RequestJWTToken(ctx context.Context, client proto.DaemonServiceClient, stdout, stderr io.Writer, useCache bool, hint string, openBrowser func(string) error) (string, error) {
-	req := &proto.RequestJWTAuthRequest{}
+	// the ssh client runs in the user's session, the daemon does not: tell it what we can see
+	req := &proto.RequestJWTAuthRequest{HasGraphicalSession: util.HasGraphicalSession()}
 	if hint != "" {
 		req.Hint = &hint
 	}
@@ -193,4 +195,3 @@ func buildAddressList(hostname string, remote net.Addr) []string {
 	}
 	return addresses
 }
-

client/ssh/server/command_execution_windows.go

@@ -243,7 +243,7 @@ func (s *Server) setUserEnvironmentVariables(envMap map[string]string, userProfi
 
 // prepareCommandEnv prepares environment variables for command execution on Windows
 func (s *Server) prepareCommandEnv(logger *log.Entry, localUser *user.User, session ssh.Session) []string {
-	username, domain := s.parseUsername(localUser.Username)
+	username, domain := parseUsername(localUser.Username)
 	userEnv, err := s.getUserEnvironment(logger, username, domain)
 	if err != nil {
 		log.Debugf("failed to get user environment for %s\\%s, using fallback: %v", domain, username, err)
@@ -383,7 +383,7 @@ func (s *Server) executeCommandWithPty(logger *log.Entry, session ssh.Session, _
 		return false
 	}
 
-	username, domain := s.parseUsername(localUser.Username)
+	username, domain := parseUsername(localUser.Username)
 	shell := getUserShell(localUser.Uid)
 
 	req := PtyExecutionRequest{

client/ssh/server/port_forwarding.go

@@ -133,7 +133,12 @@ func (s *Server) checkPrivilegedPortAccess(forwardType string, port uint32, resu
 		return nil
 	}
 
-	if result.User != nil && isPrivilegedUsername(result.User.Username) {
+	// Only uid 0 may bind below the threshold, which is the kernel's own rule and
+	// is asked directly rather than through isPrivilegedOrUnknown: that helper
+	// reports an account it cannot evaluate as privileged, which is safe for a
+	// refusal and unsafe for a grant such as this one. Windows has returned
+	// above, so Uid here is a Unix uid and never a SID.
+	if result.User != nil && result.User.Uid == "0" {
 		return nil
 	}
 

client/ssh/server/privileges_other.go

@@ -0,0 +1,16 @@
+//go:build !windows
+
+package server
+
+// isProcessElevated is only meaningful on Windows; other platforms use the
+// effective UID check in isCurrentProcessPrivileged.
+func isProcessElevated() bool {
+	return false
+}
+
+// isWindowsAccountPrivilegedOrUnknown is only reachable on Windows. Report
+// privileged on other platforms so a caller refusing privileged accounts fails
+// closed.
+func isWindowsAccountPrivilegedOrUnknown(string) bool {
+	return true
+}

client/ssh/server/privileges_windows.go

@@ -0,0 +1,228 @@
+//go:build windows
+
+package server
+
+import (
+	"fmt"
+	"strings"
+	"unsafe"
+
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/sys/windows"
+)
+
+var (
+	netapi32                  = windows.NewLazySystemDLL("netapi32.dll")
+	procNetUserGetLocalGroups = netapi32.NewProc("NetUserGetLocalGroups")
+)
+
+const (
+	// lgIncludeIndirect makes NetUserGetLocalGroups also return local groups
+	// the user belongs to through a global group.
+	lgIncludeIndirect  = 0x1
+	maxPreferredLength = 0xFFFFFFFF
+)
+
+// localGroupUsersInfo0 mirrors LOCALGROUP_USERS_INFO_0.
+type localGroupUsersInfo0 struct {
+	name *uint16
+}
+
+// isProcessElevated reports whether the current process token is elevated
+// (TokenElevation): true for elevated administrators, the built-in
+// Administrator, administrators with UAC disabled, and SYSTEM; false for
+// standard users and administrators running with a UAC-filtered token.
+func isProcessElevated() bool {
+	return windows.GetCurrentProcessToken().IsElevated()
+}
+
+// isWindowsAccountPrivilegedOrUnknown reports whether the account is privileged
+// on this machine: a well-known service account, a built-in Administrator
+// (RID 500), or a member of the local Administrators group, directly or through
+// nested groups.
+//
+// An account whose privilege cannot be determined counts as privileged, which
+// is why the name says "or unknown". That is fail-closed for a caller that
+// refuses privileged accounts, and fail-open for a caller that grants something
+// to them, so only the former may use this.
+func isWindowsAccountPrivilegedOrUnknown(username string) bool {
+	sid, _, _, err := windows.LookupSID("", username)
+	if err != nil {
+		log.Warnf("privilege check: SID lookup for %q failed, treating as privileged: %v", username, err)
+		return true
+	}
+
+	if isPrivilegedUserSID(sid) {
+		return true
+	}
+
+	member, err := isLocalAdminsMember(username)
+	if err != nil {
+		log.Warnf("privilege check: cannot determine Administrators membership for %q, treating as privileged: %v", username, err)
+		return true
+	}
+	return member
+}
+
+// isPrivilegedUserSID reports whether the SID itself identifies a privileged
+// principal, without consulting group membership.
+func isPrivilegedUserSID(sid *windows.SID) bool {
+	wellKnown := []windows.WELL_KNOWN_SID_TYPE{
+		windows.WinLocalSystemSid,
+		windows.WinLocalServiceSid,
+		windows.WinNetworkServiceSid,
+		windows.WinBuiltinAdministratorsSid,
+	}
+	for _, sidType := range wellKnown {
+		if sid.IsWellKnown(sidType) {
+			return true
+		}
+	}
+	return isBuiltinAdministratorSID(sid)
+}
+
+// isBuiltinAdministratorSID reports whether the SID is a machine or domain
+// built-in Administrator account (S-1-5-21-...-500). RID 500 is reserved for
+// that account; it can be renamed but cannot be removed from the
+// Administrators group.
+func isBuiltinAdministratorSID(sid *windows.SID) bool {
+	if sid.IdentifierAuthority() != windows.SECURITY_NT_AUTHORITY {
+		return false
+	}
+	count := sid.SubAuthorityCount()
+	if count < 2 || sid.SubAuthority(0) != 21 {
+		return false
+	}
+	return sid.SubAuthority(uint32(count-1)) == 500
+}
+
+// isLocalAdminsMember reports whether the account is a member of the local
+// Administrators group.
+//
+// Local accounts are checked against the local SAM, which is authoritative for
+// them and, unlike a token, cannot under-report: UAC filters the tokens of
+// local administrators, and a filtered token carries Administrators as
+// deny-only, which a membership check on the token would read as "not a
+// member". Domain accounts are exempt from that filtering, so for them an S4U
+// token is preferred because its group list is LSA's transitive expansion and
+// therefore covers nested and universal groups plus the machine's own local
+// groups. NetUserGetLocalGroups expands only one global-group hop but needs no
+// logon, so it serves as the fallback when no token can be obtained.
+func isLocalAdminsMember(username string) (bool, error) {
+	adminSid, err := windows.CreateWellKnownSid(windows.WinBuiltinAdministratorsSid)
+	if err != nil {
+		return false, fmt.Errorf("create Administrators SID: %w", err)
+	}
+
+	account, domain := parseUsername(username)
+	if NewPrivilegeDropper().isLocalUser(domain) {
+		return localGroupsContainSID(account, adminSid)
+	}
+
+	member, s4uErr := s4uTokenIsMember(account, domain, adminSid)
+	if s4uErr == nil {
+		return member, nil
+	}
+	log.Debugf("privilege check: S4U membership check for %q failed, falling back to local group enumeration: %v", username, s4uErr)
+
+	member, err = localGroupsContainSID(buildUserCpn(account, domain), adminSid)
+	if err != nil {
+		return false, fmt.Errorf("S4U check: %w; local group enumeration: %w", s4uErr, err)
+	}
+	return member, nil
+}
+
+// s4uTokenIsMember obtains an S4U token for the account and checks whether the
+// given SID is enabled in it.
+func s4uTokenIsMember(account, domain string, sid *windows.SID) (bool, error) {
+	token, err := generateS4UUserToken(log.NewEntry(log.StandardLogger()), account, domain)
+	if err != nil {
+		return false, err
+	}
+	defer func() {
+		if err := windows.CloseHandle(token); err != nil {
+			log.Debugf("close S4U token: %v", err)
+		}
+	}()
+	return windows.Token(token).IsMember(sid)
+}
+
+// localGroupsContainSID reports whether the wanted group is among the local
+// groups the account belongs to, directly or through a global group.
+//
+// The wanted SID is resolved to its group name once and compared against the
+// enumerated names. Well-known SIDs resolve from a static table, so that lookup
+// needs no domain controller, and it keeps the comparison correct for a renamed
+// or localized group because both sides then carry the new name. Resolving each
+// enumerated name back to a SID instead would add a lookup per group that can
+// block until it times out while a domain controller is unreachable, and cannot
+// change the outcome: the names enumerated here are local groups of this
+// machine, whose names are unique, so a name match identifies the group.
+//
+// A failure to resolve the wanted SID is returned rather than reported as
+// "not a member", so a privilege check built on this fails closed.
+func localGroupsContainSID(username string, want *windows.SID) (bool, error) {
+	wantName, _, _, err := want.LookupAccount("")
+	if err != nil {
+		return false, fmt.Errorf("resolve group SID %s to a name: %w", want, err)
+	}
+
+	groups, err := netUserGetLocalGroups(username)
+	if err != nil {
+		return false, err
+	}
+
+	for _, group := range groups {
+		if strings.EqualFold(group, wantName) {
+			return true, nil
+		}
+	}
+	return false, nil
+}
+
+// netUserGetLocalGroups returns the names of the local groups the account is a
+// member of, including indirect membership through global groups.
+func netUserGetLocalGroups(username string) ([]string, error) {
+	name16, err := windows.UTF16PtrFromString(username)
+	if err != nil {
+		return nil, fmt.Errorf("convert username: %w", err)
+	}
+
+	var buf *byte
+	var entriesRead, totalEntries uint32
+	status, _, _ := procNetUserGetLocalGroups.Call(
+		0, // local server
+		uintptr(unsafe.Pointer(name16)),
+		0, // level 0: LOCALGROUP_USERS_INFO_0
+		lgIncludeIndirect,
+		uintptr(unsafe.Pointer(&buf)),
+		maxPreferredLength,
+		uintptr(unsafe.Pointer(&entriesRead)),
+		uintptr(unsafe.Pointer(&totalEntries)),
+	)
+	if status != 0 {
+		return nil, fmt.Errorf("NetUserGetLocalGroups for %q: status %d", username, status)
+	}
+	if buf == nil {
+		return nil, nil
+	}
+	defer func() {
+		if err := windows.NetApiBufferFree(buf); err != nil {
+			log.Debugf("free NetApi buffer: %v", err)
+		}
+	}()
+
+	// MAX_PREFERRED_LENGTH makes the API allocate as much as it needs, so a
+	// short read is not expected. Report it rather than silently returning a
+	// subset of the account's groups.
+	if entriesRead != totalEntries {
+		return nil, fmt.Errorf("NetUserGetLocalGroups for %q returned %d of %d groups", username, entriesRead, totalEntries)
+	}
+
+	entries := unsafe.Slice((*localGroupUsersInfo0)(unsafe.Pointer(buf)), entriesRead)
+	groups := make([]string, 0, entriesRead)
+	for _, entry := range entries {
+		groups = append(groups, windows.UTF16PtrToString(entry.name))
+	}
+	return groups, nil
+}

client/ssh/server/privileges_windows_test.go

@@ -0,0 +1,293 @@
+//go:build windows
+
+package server
+
+import (
+	"os/user"
+	"testing"
+	"unsafe"
+
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/sys/windows"
+)
+
+// filterNormalAccount limits NetUserEnum to normal user accounts.
+const filterNormalAccount = 0x2
+
+// TOKEN_ELEVATION_TYPE values.
+const (
+	tokenElevationTypeDefault = 1
+	tokenElevationTypeFull    = 2
+	tokenElevationTypeLimited = 3
+)
+
+// tokenElevationType reads TokenElevationType from a token.
+func tokenElevationType(token windows.Token) (uint32, error) {
+	var elevationType, returnedLen uint32
+	err := windows.GetTokenInformation(token, windows.TokenElevationType,
+		(*byte)(unsafe.Pointer(&elevationType)), uint32(unsafe.Sizeof(elevationType)), &returnedLen)
+	if err != nil {
+		return 0, err
+	}
+	return elevationType, nil
+}
+
+// userInfo0 mirrors USER_INFO_0.
+type userInfo0 struct {
+	name *uint16
+}
+
+func mustParseSID(t *testing.T, s string) *windows.SID {
+	t.Helper()
+	sid, err := windows.StringToSid(s)
+	require.NoError(t, err, "parse SID %s", s)
+	return sid
+}
+
+// localAccountNames returns the names of the local user accounts.
+func localAccountNames(t *testing.T) []string {
+	t.Helper()
+
+	var buf *byte
+	var entriesRead, totalEntries, resume uint32
+	err := windows.NetUserEnum(nil, 0, filterNormalAccount, &buf, maxPreferredLength,
+		&entriesRead, &totalEntries, &resume)
+	require.NoError(t, err, "enumerate local users")
+	t.Cleanup(func() {
+		require.NoError(t, windows.NetApiBufferFree(buf), "free NetApi buffer")
+	})
+
+	entries := unsafe.Slice((*userInfo0)(unsafe.Pointer(buf)), entriesRead)
+	names := make([]string, 0, entriesRead)
+	for _, entry := range entries {
+		names = append(names, windows.UTF16PtrToString(entry.name))
+	}
+	return names
+}
+
+// localAccountNameByRID returns the name of the local account carrying the
+// given RID. Accounts such as Administrator and Guest can be renamed and are
+// localized, so tests must not name them literally.
+func localAccountNameByRID(t *testing.T, rid uint32) string {
+	t.Helper()
+
+	for _, name := range localAccountNames(t) {
+		sid, _, _, err := windows.LookupSID("", name)
+		if err != nil {
+			continue
+		}
+		if sid.IdentifierAuthority() != windows.SECURITY_NT_AUTHORITY {
+			continue
+		}
+		count := sid.SubAuthorityCount()
+		if count < 2 || sid.SubAuthority(0) != 21 {
+			continue
+		}
+		if sid.SubAuthority(uint32(count-1)) == rid {
+			return name
+		}
+	}
+
+	t.Fatalf("no local account with RID %d", rid)
+	return ""
+}
+
+// wellKnownAccountName resolves a well-known SID to the qualified account name
+// the local system uses for it, which is localized.
+func wellKnownAccountName(t *testing.T, sidType windows.WELL_KNOWN_SID_TYPE) string {
+	t.Helper()
+
+	sid, err := windows.CreateWellKnownSid(sidType)
+	require.NoError(t, err, "create well-known SID")
+	name, domain, _, err := sid.LookupAccount("")
+	require.NoError(t, err, "resolve %s to an account name", sid)
+	if domain == "" {
+		return name
+	}
+	return domain + `\` + name
+}
+
+func TestIsBuiltinAdministratorSID(t *testing.T) {
+	tests := []struct {
+		name string
+		sid  string
+		want bool
+	}{
+		{"machine_administrator", "S-1-5-21-1111111111-2222222222-3333333333-500", true},
+		{"domain_administrator", "S-1-5-21-3390233681-4087452608-412898826-500", true},
+		{"regular_user", "S-1-5-21-1111111111-2222222222-3333333333-1001", false},
+		{"guest_account", "S-1-5-21-1111111111-2222222222-3333333333-501", false},
+		{"domain_admins_group", "S-1-5-21-1111111111-2222222222-3333333333-512", false},
+		{"system", "S-1-5-18", false},
+		{"administrators_group", "S-1-5-32-544", false},
+		{"non_nt_authority", "S-1-1-0", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := isBuiltinAdministratorSID(mustParseSID(t, tt.sid))
+			assert.Equal(t, tt.want, result, "RID 500 detection for %s", tt.sid)
+		})
+	}
+}
+
+func TestIsPrivilegedUserSID(t *testing.T) {
+	tests := []struct {
+		name string
+		sid  string
+		want bool
+	}{
+		{"local_system", "S-1-5-18", true},
+		{"local_service", "S-1-5-19", true},
+		{"network_service", "S-1-5-20", true},
+		{"administrators_group", "S-1-5-32-544", true},
+		{"builtin_administrator", "S-1-5-21-1111111111-2222222222-3333333333-500", true},
+		{"regular_user", "S-1-5-21-1111111111-2222222222-3333333333-1001", false},
+		{"users_group", "S-1-5-32-545", false},
+		{"everyone", "S-1-1-0", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := isPrivilegedUserSID(mustParseSID(t, tt.sid))
+			assert.Equal(t, tt.want, result, "SID privilege classification for %s", tt.sid)
+		})
+	}
+}
+
+func TestIsWindowsAccountPrivilegedOrUnknown(t *testing.T) {
+	tests := []struct {
+		name     string
+		username string
+		want     bool
+	}{
+		{"system", wellKnownAccountName(t, windows.WinLocalSystemSid), true},
+		{"local_service", wellKnownAccountName(t, windows.WinLocalServiceSid), true},
+		{"network_service", wellKnownAccountName(t, windows.WinNetworkServiceSid), true},
+		{"administrators_group", wellKnownAccountName(t, windows.WinBuiltinAdministratorsSid), true},
+		// The built-in Administrator (RID 500) and Guest (RID 501) accounts
+		// exist on every Windows installation, though they may be disabled.
+		{"builtin_administrator", localAccountNameByRID(t, 500), true},
+		{"guest", localAccountNameByRID(t, 501), false},
+		// Unresolvable accounts fail closed.
+		{"nonexistent_user", "netbird-no-such-user", true},
+		{"empty_username", "", true},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			result := isWindowsAccountPrivilegedOrUnknown(tt.username)
+			assert.Equal(t, tt.want, result, "account privilege classification for %q", tt.username)
+		})
+	}
+}
+
+func TestIsProcessElevated(t *testing.T) {
+	elevated := isProcessElevated()
+
+	// TokenElevationType is a second, independent view of the same token:
+	// Full means elevated and Limited means a filtered administrator, while
+	// Default covers both a standard user and an administrator with no linked
+	// token (UAC off, the built-in Administrator, SYSTEM), so it implies nothing.
+	elevationType, err := tokenElevationType(windows.GetCurrentProcessToken())
+	require.NoError(t, err, "read token elevation type")
+
+	adminSid, err := windows.CreateWellKnownSid(windows.WinBuiltinAdministratorsSid)
+	require.NoError(t, err, "create Administrators SID")
+
+	// Token(0) makes CheckTokenMembership evaluate the caller's own token. It
+	// counts only enabled SIDs, so a filtered administrator reports false here.
+	member, err := windows.Token(0).IsMember(adminSid)
+	require.NoError(t, err, "check own Administrators membership")
+
+	t.Logf("elevated=%v elevationType=%d memberOfAdministrators=%v", elevated, elevationType, member)
+
+	switch elevationType {
+	case tokenElevationTypeFull:
+		assert.True(t, elevated, "a token of elevation type Full must report elevated")
+	case tokenElevationTypeLimited:
+		assert.False(t, elevated, "a filtered administrator token must not report elevated")
+	}
+
+	// Administrators enabled in the token means the token wields administrative
+	// rights, which is what elevation reports.
+	if member {
+		assert.True(t, elevated, "token with enabled Administrators membership must report elevated")
+	}
+}
+
+// TestS4UMembershipAgreesWithLocalGroups exercises the S4U token path used
+// for domain accounts. S4U logons need the TCB privilege, so the test runs
+// only as SYSTEM (which is how CI executes the suite). For local accounts the
+// token's Administrators membership must agree with the SAM enumeration.
+func TestS4UMembershipAgreesWithLocalGroups(t *testing.T) {
+	system, err := windows.CreateWellKnownSid(windows.WinLocalSystemSid)
+	require.NoError(t, err, "create SYSTEM SID")
+	current, err := user.Current()
+	require.NoError(t, err, "get current user")
+	if current.Uid != system.String() {
+		t.Skipf("S4U logon requires SYSTEM (running as %s)", current.Username)
+	}
+
+	adminSid, err := windows.CreateWellKnownSid(windows.WinBuiltinAdministratorsSid)
+	require.NoError(t, err, "create Administrators SID")
+
+	checked := 0
+	for _, name := range localAccountNames(t) {
+		viaToken, err := s4uTokenIsMember(name, ".", adminSid)
+		if err != nil {
+			// Disabled or logon-restricted accounts cannot get an S4U logon.
+			t.Logf("skipping %s: %v", name, err)
+			continue
+		}
+		viaSAM, err := localGroupsContainSID(name, adminSid)
+		require.NoError(t, err, "enumerate local groups for %s", name)
+
+		assert.Equal(t, viaSAM, viaToken, "S4U token and SAM enumeration must agree on Administrators membership for %s", name)
+		checked++
+	}
+	// Ineligible accounts are skipped, so without this the test could report
+	// success while comparing nothing at all.
+	require.Positive(t, checked, "no local account completed an S4U logon, so nothing was compared")
+	t.Logf("checked %d local accounts via S4U", checked)
+}
+
+// TestLocalGroupsContainSID_Administrator checks the positive case against the
+// built-in Administrator, a member of Administrators on every installation.
+func TestLocalGroupsContainSID_Administrator(t *testing.T) {
+	adminSid, err := windows.CreateWellKnownSid(windows.WinBuiltinAdministratorsSid)
+	require.NoError(t, err, "create Administrators SID")
+
+	administrator := localAccountNameByRID(t, 500)
+	member, err := localGroupsContainSID(administrator, adminSid)
+	require.NoError(t, err, "enumerate local groups for %s", administrator)
+	assert.True(t, member, "%s is a member of the Administrators group", administrator)
+}
+
+// TestLocalGroupsContainSID_UnresolvableGroupFailsClosed covers a wanted SID
+// that resolves to no group: the error must surface rather than being reported
+// as "not a member", so the privilege check treats the account as privileged.
+func TestLocalGroupsContainSID_UnresolvableGroupFailsClosed(t *testing.T) {
+	unknown := mustParseSID(t, "S-1-5-21-1111111111-2222222222-3333333333-4444")
+
+	_, err := localGroupsContainSID(localAccountNameByRID(t, 500), unknown)
+	require.Error(t, err, "must report an error when the wanted group cannot be identified")
+}
+
+func TestLocalGroupsContainSID_Guest(t *testing.T) {
+	guestsSid, err := windows.CreateWellKnownSid(windows.WinBuiltinGuestsSid)
+	require.NoError(t, err, "create Guests SID")
+	adminsSid, err := windows.CreateWellKnownSid(windows.WinBuiltinAdministratorsSid)
+	require.NoError(t, err, "create Administrators SID")
+
+	guest := localAccountNameByRID(t, 501)
+
+	inGuests, err := localGroupsContainSID(guest, guestsSid)
+	require.NoError(t, err, "enumerate local groups for %s", guest)
+	assert.True(t, inGuests, "%s is a member of the Guests group", guest)
+
+	inAdmins, err := localGroupsContainSID(guest, adminsSid)
+	require.NoError(t, err, "enumerate local groups for %s", guest)
+	assert.False(t, inAdmins, "%s is not a member of the Administrators group", guest)
+}

client/ssh/server/server_config_test.go

@@ -239,6 +239,7 @@ func TestServer_PrivilegedPortAccess(t *testing.T) {
 		forwardType   string
 		port          uint32
 		username      string
+		uid           string
 		expectError   bool
 		errorMsg      string
 		skipOnWindows bool
@@ -248,6 +249,7 @@ func TestServer_PrivilegedPortAccess(t *testing.T) {
 			forwardType:   "remote",
 			port:          80,
 			username:      "testuser",
+			uid:           "1000",
 			expectError:   true,
 			errorMsg:      "cannot bind to privileged port",
 			skipOnWindows: true,
@@ -257,6 +259,7 @@ func TestServer_PrivilegedPortAccess(t *testing.T) {
 			forwardType:   "tcpip-forward",
 			port:          443,
 			username:      "testuser",
+			uid:           "1000",
 			expectError:   true,
 			errorMsg:      "cannot bind to privileged port",
 			skipOnWindows: true,
@@ -266,6 +269,7 @@ func TestServer_PrivilegedPortAccess(t *testing.T) {
 			forwardType: "remote",
 			port:        8080,
 			username:    "testuser",
+			uid:         "1000",
 			expectError: false,
 		},
 		{
@@ -273,6 +277,7 @@ func TestServer_PrivilegedPortAccess(t *testing.T) {
 			forwardType: "remote",
 			port:        0,
 			username:    "testuser",
+			uid:         "1000",
 			expectError: false,
 		},
 		{
@@ -280,13 +285,35 @@ func TestServer_PrivilegedPortAccess(t *testing.T) {
 			forwardType: "remote",
 			port:        22,
 			username:    "root",
+			uid:         "0",
 			expectError: false,
 		},
+		{
+			// Only uid 0 is privileged, whatever the account is called.
+			name:          "uid 0 under another name may bind a privileged port",
+			forwardType:   "remote",
+			port:          22,
+			username:      "toor",
+			uid:           "0",
+			expectError:   false,
+			skipOnWindows: true,
+		},
+		{
+			name:          "account named root without uid 0 may not",
+			forwardType:   "remote",
+			port:          22,
+			username:      "root",
+			uid:           "1000",
+			expectError:   true,
+			errorMsg:      "cannot bind to privileged port",
+			skipOnWindows: true,
+		},
 		{
 			name:        "local forward privileged port allowed for non-root",
 			forwardType: "local",
 			port:        80,
 			username:    "testuser",
+			uid:         "1000",
 			expectError: false,
 		},
 	}
@@ -299,7 +326,7 @@ func TestServer_PrivilegedPortAccess(t *testing.T) {
 
 			result := PrivilegeCheckResult{
 				Allowed: true,
-				User:    &user.User{Username: tt.username},
+				User:    &user.User{Username: tt.username, Uid: tt.uid},
 			}
 
 			err := server.checkPrivilegedPortAccess(tt.forwardType, tt.port, result)
@@ -420,6 +447,13 @@ func TestServer_PortConflictHandling(t *testing.T) {
 
 func TestServer_IsPrivilegedUser(t *testing.T) {
 
+	// Windows classification depends on account SIDs and group membership, and
+	// the accounts involved carry localized, renameable names. It is covered by
+	// TestIsWindowsAccountPrivileged, which resolves them from well-known SIDs.
+	if runtime.GOOS == "windows" {
+		t.Skip("covered by TestIsWindowsAccountPrivileged")
+	}
+
 	tests := []struct {
 		username    string
 		expected    bool
@@ -440,44 +474,16 @@ func TestServer_IsPrivilegedUser(t *testing.T) {
 			expected:    false,
 			description: "empty username should not be privileged",
 		},
-	}
-
-	// Add Windows-specific tests
-	if runtime.GOOS == "windows" {
-		tests = append(tests, []struct {
-			username    string
-			expected    bool
-			description string
-		}{
-			{
-				username:    "Administrator",
-				expected:    true,
-				description: "Administrator should be considered privileged on Windows",
-			},
-			{
-				username:    "administrator",
-				expected:    true,
-				description: "administrator should be considered privileged on Windows (case insensitive)",
-			},
-		}...)
-	} else {
-		// On non-Windows systems, Administrator should not be privileged
-		tests = append(tests, []struct {
-			username    string
-			expected    bool
-			description string
-		}{
-			{
-				username:    "Administrator",
-				expected:    false,
-				description: "Administrator should not be privileged on non-Windows systems",
-			},
-		}...)
+		{
+			username:    "Administrator",
+			expected:    false,
+			description: "Administrator should not be privileged on non-Windows systems",
+		},
 	}
 
 	for _, tt := range tests {
 		t.Run(tt.description, func(t *testing.T) {
-			result := isPrivilegedUsername(tt.username)
+			result := isPrivilegedOrUnknown(tt.username)
 			assert.Equal(t, tt.expected, result, tt.description)
 		})
 	}

client/ssh/server/sftp_windows.go

@@ -17,7 +17,7 @@ import (
 // createSftpCommand creates a Windows SFTP command with user switching.
 // The caller must close the returned token handle after starting the process.
 func (s *Server) createSftpCommand(targetUser *user.User, sess ssh.Session) (*exec.Cmd, windows.Token, error) {
-	username, domain := s.parseUsername(targetUser.Username)
+	username, domain := parseUsername(targetUser.Username)
 
 	netbirdPath, err := os.Executable()
 	if err != nil {

client/ssh/server/user_utils.go

@@ -16,11 +16,6 @@ var (
 	ErrPrivilegedUserSwitch = errors.New("cannot switch to privileged user - current user lacks required privileges")
 )
 
-// isPlatformUnix returns true for Unix-like platforms (Linux, macOS, etc.)
-func isPlatformUnix() bool {
-	return getCurrentOS() != "windows"
-}
-
 // Dependency injection variables for testing - allows mocking dynamic runtime checks
 var (
 	getCurrentUser         = currentUserWithGetent
@@ -29,6 +24,9 @@ var (
 	getIsProcessPrivileged = isCurrentProcessPrivileged
 
 	getEuid = os.Geteuid
+
+	getProcessElevated                   = isProcessElevated
+	getWindowsAccountPrivilegedOrUnknown = isWindowsAccountPrivilegedOrUnknown
 )
 
 const (
@@ -65,6 +63,13 @@ type PrivilegeCheckResult struct {
 	RequiresUserSwitching bool
 }
 
+// privilegeCheckContext holds all context needed for privilege checking
+type privilegeCheckContext struct {
+	currentUser           *user.User
+	currentUserPrivileged bool
+	allowRoot             bool
+}
+
 // CheckPrivileges performs comprehensive privilege checking for all SSH features.
 // This is the single source of truth for privilege decisions across the SSH server.
 func (s *Server) CheckPrivileges(req PrivilegeCheckRequest) PrivilegeCheckResult {
@@ -75,7 +80,7 @@ func (s *Server) CheckPrivileges(req PrivilegeCheckRequest) PrivilegeCheckResult
 
 	// Handle empty username case - but still check root access controls
 	if req.RequestedUsername == "" {
-		if isPrivilegedUsername(context.currentUser.Username) && !context.allowRoot {
+		if isPrivilegedOrUnknown(context.currentUser.Username) && !context.allowRoot {
 			return PrivilegeCheckResult{
 				Allowed: false,
 				Error:   &PrivilegedUserError{Username: context.currentUser.Username},
@@ -135,7 +140,7 @@ func (s *Server) checkUserRequest(ctx *privilegeCheckContext, req PrivilegeCheck
 
 	needsUserSwitching := !isSameResolvedUser(resolvedUser, ctx.currentUser)
 
-	if isPrivilegedUsername(resolvedUser.Username) && !ctx.allowRoot {
+	if isPrivilegedOrUnknown(resolvedUser.Username) && !ctx.allowRoot {
 		return PrivilegeCheckResult{
 			Allowed:               false,
 			Error:                 &PrivilegedUserError{Username: resolvedUser.Username},
@@ -175,6 +180,42 @@ func (s *Server) resolveRequestedUser(requestedUsername string) (*user.User, err
 	return u, nil
 }
 
+// SetAllowRootLogin configures root login access
+func (s *Server) SetAllowRootLogin(allow bool) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.allowRootLogin = allow
+}
+
+// userNameLookup performs user lookup with root login permission check
+func (s *Server) userNameLookup(username string) (*user.User, error) {
+	result, err := s.userPrivilegeCheck(username)
+	if err != nil {
+		return nil, err
+	}
+	return result.User, nil
+}
+
+// userPrivilegeCheck performs user lookup with full privilege check result
+func (s *Server) userPrivilegeCheck(username string) (PrivilegeCheckResult, error) {
+	result := s.CheckPrivileges(PrivilegeCheckRequest{
+		RequestedUsername:         username,
+		FeatureSupportsUserSwitch: true,
+		FeatureName:               FeatureSSHLogin,
+	})
+
+	if !result.Allowed {
+		return result, result.Error
+	}
+
+	return result, nil
+}
+
+// isPlatformUnix returns true for Unix-like platforms (Linux, macOS, etc.)
+func isPlatformUnix() bool {
+	return getCurrentOS() != "windows"
+}
+
 // isSameResolvedUser compares two resolved user identities
 func isSameResolvedUser(user1, user2 *user.User) bool {
 	if user1 == nil || user2 == nil {
@@ -183,13 +224,6 @@ func isSameResolvedUser(user1, user2 *user.User) bool {
 	return user1.Uid == user2.Uid
 }
 
-// privilegeCheckContext holds all context needed for privilege checking
-type privilegeCheckContext struct {
-	currentUser           *user.User
-	currentUserPrivileged bool
-	allowRoot             bool
-}
-
 // isSameUser checks if two usernames refer to the same user
 // SECURITY: This function must be conservative - it should only return true
 // when we're certain both usernames refer to the exact same user identity
@@ -253,159 +287,30 @@ func isWindowsSameUser(requestedUsername, currentUsername string) bool {
 	return strings.EqualFold(reqDomain, curDomain)
 }
 
-// SetAllowRootLogin configures root login access
-func (s *Server) SetAllowRootLogin(allow bool) {
-	s.mu.Lock()
-	defer s.mu.Unlock()
-	s.allowRootLogin = allow
-}
-
-// userNameLookup performs user lookup with root login permission check
-func (s *Server) userNameLookup(username string) (*user.User, error) {
-	result := s.CheckPrivileges(PrivilegeCheckRequest{
-		RequestedUsername:         username,
-		FeatureSupportsUserSwitch: true,
-		FeatureName:               FeatureSSHLogin,
-	})
-
-	if !result.Allowed {
-		return nil, result.Error
-	}
-
-	return result.User, nil
-}
-
-// userPrivilegeCheck performs user lookup with full privilege check result
-func (s *Server) userPrivilegeCheck(username string) (PrivilegeCheckResult, error) {
-	result := s.CheckPrivileges(PrivilegeCheckRequest{
-		RequestedUsername:         username,
-		FeatureSupportsUserSwitch: true,
-		FeatureName:               FeatureSSHLogin,
-	})
-
-	if !result.Allowed {
-		return result, result.Error
-	}
-
-	return result, nil
-}
-
-// isPrivilegedUsername checks if the given username represents a privileged user across platforms.
-// On Unix: root
-// On Windows: Administrator, SYSTEM (case-insensitive)
-// Handles domain-qualified usernames like "DOMAIN\Administrator" or "user@domain.com"
-func isPrivilegedUsername(username string) bool {
+// isPrivilegedOrUnknown reports whether the given username represents a
+// privileged user, or on Windows an account whose privilege could not be
+// determined.
+// On Unix: root.
+// On Windows: well-known service accounts, built-in Administrator accounts,
+// and members of the local Administrators group; handles domain-qualified
+// usernames like "DOMAIN\user" or "user@domain.com". An account that cannot be
+// resolved or evaluated is reported as privileged.
+//
+// Use this to refuse privileged accounts, never to grant them anything: the
+// undetermined case is safe for a refusal and unsafe for a grant.
+func isPrivilegedOrUnknown(username string) bool {
 	if getCurrentOS() != "windows" {
 		return username == "root"
 	}
-
-	bareUsername := username
-	// Handle Windows domain format: DOMAIN\username
-	if idx := strings.LastIndex(username, `\`); idx != -1 {
-		bareUsername = username[idx+1:]
-	}
-	// Handle email-style format: username@domain.com
-	if idx := strings.Index(bareUsername, "@"); idx != -1 {
-		bareUsername = bareUsername[:idx]
-	}
-
-	return isWindowsPrivilegedUser(bareUsername)
-}
-
-// isWindowsPrivilegedUser checks if a bare username (domain already stripped) represents a Windows privileged account
-func isWindowsPrivilegedUser(bareUsername string) bool {
-	// common privileged usernames (case insensitive)
-	privilegedNames := []string{
-		"administrator",
-		"admin",
-		"root",
-		"system",
-		"localsystem",
-		"networkservice",
-		"localservice",
-	}
-
-	usernameLower := strings.ToLower(bareUsername)
-	for _, privilegedName := range privilegedNames {
-		if usernameLower == privilegedName {
-			return true
-		}
-	}
-
-	// computer accounts (ending with $) are not privileged by themselves
-	// They only gain privileges through group membership or specific SIDs
-
-	if targetUser, err := lookupUser(bareUsername); err == nil {
-		return isWindowsPrivilegedSID(targetUser.Uid)
-	}
-
-	return false
-}
-
-// isWindowsPrivilegedSID checks if a Windows SID represents a privileged account
-func isWindowsPrivilegedSID(sid string) bool {
-	privilegedSIDs := []string{
-		"S-1-5-18",     // Local System (SYSTEM)
-		"S-1-5-19",     // Local Service (NT AUTHORITY\LOCAL SERVICE)
-		"S-1-5-20",     // Network Service (NT AUTHORITY\NETWORK SERVICE)
-		"S-1-5-32-544", // Administrators group (BUILTIN\Administrators)
-		"S-1-5-500",    // Built-in Administrator account (local machine RID 500)
-	}
-
-	for _, privilegedSID := range privilegedSIDs {
-		if sid == privilegedSID {
-			return true
-		}
-	}
-
-	// Check for domain administrator accounts (RID 500 in any domain)
-	// Format: S-1-5-21-domain-domain-domain-500
-	// This is reliable as RID 500 is reserved for the domain Administrator account
-	if strings.HasPrefix(sid, "S-1-5-21-") && strings.HasSuffix(sid, "-500") {
-		return true
-	}
-
-	// Check for other well-known privileged RIDs in domain contexts
-	// RID 512 = Domain Admins group, RID 516 = Domain Controllers group
-	if strings.HasPrefix(sid, "S-1-5-21-") {
-		if strings.HasSuffix(sid, "-512") || // Domain Admins group
-			strings.HasSuffix(sid, "-516") || // Domain Controllers group
-			strings.HasSuffix(sid, "-519") { // Enterprise Admins group
-			return true
-		}
-	}
-
-	return false
+	return getWindowsAccountPrivilegedOrUnknown(username)
 }
 
 // isCurrentProcessPrivileged checks if the current process is running with elevated privileges.
 // On Unix systems, this means running as root (UID 0).
-// On Windows, this means running as Administrator or SYSTEM.
+// On Windows, this means the process token is elevated (administrators, SYSTEM).
 func isCurrentProcessPrivileged() bool {
 	if getCurrentOS() == "windows" {
-		return isWindowsElevated()
+		return getProcessElevated()
 	}
 	return getEuid() == 0
 }
-
-// isWindowsElevated checks if the current process is running with elevated privileges on Windows
-func isWindowsElevated() bool {
-	currentUser, err := getCurrentUser()
-	if err != nil {
-		log.Errorf("failed to get current user for privilege check, assuming non-privileged: %v", err)
-		return false
-	}
-
-	if isWindowsPrivilegedSID(currentUser.Uid) {
-		log.Debugf("Windows user switching supported: running as privileged SID %s", currentUser.Uid)
-		return true
-	}
-
-	if isPrivilegedUsername(currentUser.Username) {
-		log.Debugf("Windows user switching supported: running as privileged username %s", currentUser.Username)
-		return true
-	}
-
-	log.Debugf("Windows user switching not supported: not running as privileged user (current: %s)", currentUser.Uid)
-	return false
-}

client/ssh/server/user_utils_test.go

@@ -4,6 +4,7 @@ import (
 	"errors"
 	"os/user"
 	"runtime"
+	"strings"
 	"testing"
 
 	"github.com/stretchr/testify/assert"
@@ -27,8 +28,8 @@ func setupTestDependencies(currentUser *user.User, currentUserErr error, os stri
 	originalLookupUser := lookupUser
 	originalGetCurrentOS := getCurrentOS
 	originalGetEuid := getEuid
-
-	// Reset caches to ensure clean test state
+	originalGetProcessElevated := getProcessElevated
+	originalGetWindowsAccountPrivilegedOrUnknown := getWindowsAccountPrivilegedOrUnknown
 
 	// Set test values - inject platform dependencies
 	getCurrentUser = func() (*user.User, error) {
@@ -53,16 +54,31 @@ func setupTestDependencies(currentUser *user.User, currentUserErr error, os stri
 		return euid
 	}
 
-	// Mock privilege detection based on the test user
-	getIsProcessPrivileged = func() bool {
+	// Simulate the Windows token elevation check based on the fixture user:
+	// the built-in Administrator (RID 500) and SYSTEM run elevated.
+	getProcessElevated = func() bool {
 		if currentUser == nil {
 			return false
 		}
-		// Check both username and SID for Windows systems
-		if os == "windows" && isWindowsPrivilegedSID(currentUser.Uid) {
+		return currentUser.Uid == "S-1-5-18" || strings.HasSuffix(currentUser.Uid, "-500")
+	}
+
+	// Simulate the Windows account classifier for the fixture accounts.
+	// "root" does not exist on Windows; the real classifier fails closed on
+	// unresolvable accounts, so it counts as privileged here too.
+	getWindowsAccountPrivilegedOrUnknown = func(username string) bool {
+		bare := username
+		if idx := strings.LastIndex(bare, `\`); idx != -1 {
+			bare = bare[idx+1:]
+		}
+		if idx := strings.Index(bare, "@"); idx != -1 {
+			bare = bare[:idx]
+		}
+		switch strings.ToLower(bare) {
+		case "administrator", "system", "root":
 			return true
 		}
-		return isPrivilegedUsername(currentUser.Username)
+		return false
 	}
 
 	// Return cleanup function
@@ -71,10 +87,8 @@ func setupTestDependencies(currentUser *user.User, currentUserErr error, os stri
 		lookupUser = originalLookupUser
 		getCurrentOS = originalGetCurrentOS
 		getEuid = originalGetEuid
-
-		getIsProcessPrivileged = isCurrentProcessPrivileged
-
-		// Reset caches after test
+		getProcessElevated = originalGetProcessElevated
+		getWindowsAccountPrivilegedOrUnknown = originalGetWindowsAccountPrivilegedOrUnknown
 	}
 }
 
@@ -421,6 +435,9 @@ func TestUsedFallback_MeansNoPrivilegeDropping(t *testing.T) {
 }
 
 func TestPrivilegedUsernameDetection(t *testing.T) {
+	// Windows classification is syscall-backed (SID resolution, group
+	// membership) and is covered by privileges_windows_test.go; here only the
+	// Unix logic and the platform dispatch are exercised.
 	tests := []struct {
 		name       string
 		username   string
@@ -432,25 +449,9 @@ func TestPrivilegedUsernameDetection(t *testing.T) {
 		{"unix_regular_user", "alice", "linux", false},
 		{"unix_root_capital", "Root", "linux", false}, // Case-sensitive
 
-		// Windows tests
+		// Windows dispatch to the (mocked) account classifier
 		{"windows_administrator", "Administrator", "windows", true},
-		{"windows_system", "SYSTEM", "windows", true},
-		{"windows_admin", "admin", "windows", true},
-		{"windows_admin_lowercase", "administrator", "windows", true}, // Case-insensitive
-		{"windows_domain_admin", "DOMAIN\\Administrator", "windows", true},
-		{"windows_email_admin", "admin@domain.com", "windows", true},
 		{"windows_regular_user", "alice", "windows", false},
-		{"windows_domain_user", "DOMAIN\\alice", "windows", false},
-		{"windows_localsystem", "localsystem", "windows", true},
-		{"windows_networkservice", "networkservice", "windows", true},
-		{"windows_localservice", "localservice", "windows", true},
-
-		// Computer accounts (these depend on current user context in real implementation)
-		{"windows_computer_account", "WIN2K19-C2$", "windows", false},      // Computer account by itself not privileged
-		{"windows_domain_computer", "DOMAIN\\COMPUTER$", "windows", false}, // Domain computer account
-
-		// Cross-platform
-		{"root_on_windows", "root", "windows", true}, // Root should be privileged everywhere
 	}
 
 	for _, tt := range tests {
@@ -459,50 +460,8 @@ func TestPrivilegedUsernameDetection(t *testing.T) {
 			cleanup := setupTestDependencies(nil, nil, tt.platform, 1000, nil, nil)
 			defer cleanup()
 
-			result := isPrivilegedUsername(tt.username)
-			assert.Equal(t, tt.privileged, result)
-		})
-	}
-}
-
-func TestWindowsPrivilegedSIDDetection(t *testing.T) {
-	tests := []struct {
-		name        string
-		sid         string
-		privileged  bool
-		description string
-	}{
-		// Well-known system accounts
-		{"system_account", "S-1-5-18", true, "Local System (SYSTEM)"},
-		{"local_service", "S-1-5-19", true, "Local Service"},
-		{"network_service", "S-1-5-20", true, "Network Service"},
-		{"administrators_group", "S-1-5-32-544", true, "Administrators group"},
-		{"builtin_administrator", "S-1-5-500", true, "Built-in Administrator"},
-
-		// Domain accounts
-		{"domain_administrator", "S-1-5-21-1234567890-1234567890-1234567890-500", true, "Domain Administrator (RID 500)"},
-		{"domain_admins_group", "S-1-5-21-1234567890-1234567890-1234567890-512", true, "Domain Admins group"},
-		{"domain_controllers_group", "S-1-5-21-1234567890-1234567890-1234567890-516", true, "Domain Controllers group"},
-		{"enterprise_admins_group", "S-1-5-21-1234567890-1234567890-1234567890-519", true, "Enterprise Admins group"},
-
-		// Regular users
-		{"regular_user", "S-1-5-21-1234567890-1234567890-1234567890-1001", false, "Regular domain user"},
-		{"another_regular_user", "S-1-5-21-1234567890-1234567890-1234567890-1234", false, "Another regular user"},
-		{"local_user", "S-1-5-21-1234567890-1234567890-1234567890-1000", false, "Local regular user"},
-
-		// Groups that are not privileged
-		{"domain_users", "S-1-5-21-1234567890-1234567890-1234567890-513", false, "Domain Users group"},
-		{"power_users", "S-1-5-32-547", false, "Power Users group"},
-
-		// Invalid SIDs
-		{"malformed_sid", "S-1-5-invalid", false, "Malformed SID"},
-		{"empty_sid", "", false, "Empty SID"},
-	}
-
-	for _, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			result := isWindowsPrivilegedSID(tt.sid)
-			assert.Equal(t, tt.privileged, result, "Failed for %s: %s", tt.description, tt.sid)
+			result := isPrivilegedOrUnknown(tt.username)
+			assert.Equal(t, tt.privileged, result, "privilege classification for %s on %s", tt.username, tt.platform)
 		})
 	}
 }

client/ssh/server/userswitching_windows.go

@@ -91,7 +91,7 @@ func validateUsernameFormat(username string) error {
 func (s *Server) createExecutorCommand(logger *log.Entry, session ssh.Session, localUser *user.User, hasPty bool) (*exec.Cmd, func(), error) {
 	logger.Debugf("creating Windows executor command for user %s (Pty: %v)", localUser.Username, hasPty)
 
-	username, _ := s.parseUsername(localUser.Username)
+	username, _ := parseUsername(localUser.Username)
 	if err := validateUsername(username); err != nil {
 		return nil, nil, fmt.Errorf("invalid username %q: %w", username, err)
 	}
@@ -102,7 +102,7 @@ func (s *Server) createExecutorCommand(logger *log.Entry, session ssh.Session, l
 // createUserSwitchCommand creates a command with Windows user switching.
 // Returns the command and a cleanup function that must be called after starting the process.
 func (s *Server) createUserSwitchCommand(logger *log.Entry, session ssh.Session, localUser *user.User) (*exec.Cmd, func(), error) {
-	username, domain := s.parseUsername(localUser.Username)
+	username, domain := parseUsername(localUser.Username)
 
 	shell := getUserShell(localUser.Uid)
 
@@ -138,7 +138,7 @@ func (s *Server) createUserSwitchCommand(logger *log.Entry, session ssh.Session,
 }
 
 // parseUsername extracts username and domain from a Windows username
-func (s *Server) parseUsername(fullUsername string) (username, domain string) {
+func parseUsername(fullUsername string) (username, domain string) {
 	// Handle DOMAIN\username format
 	if idx := strings.LastIndex(fullUsername, `\`); idx != -1 {
 		domain = fullUsername[:idx]

client/status/status.go

@@ -46,7 +46,10 @@ func ParseDaemonStatus(s string) DaemonStatus {
 
 // ConvertOptions holds parameters for ConvertToStatusOutputOverview.
 type ConvertOptions struct {
-	Anonymize            bool
+	Anonymize bool
+	// AnonymizeLevel selects how much the anonymizer redacts. Only
+	// meaningful when Anonymize is set.
+	AnonymizeLevel       anonymize.Level
 	DaemonVersion        string
 	DaemonStatus         DaemonStatus
 	StatusFilter         string
@@ -217,6 +220,7 @@ func ConvertToStatusOutputOverview(pbFullStatus *proto.FullStatus, opts ConvertO
 
 	if opts.Anonymize {
 		anonymizer := anonymize.NewAnonymizer(anonymize.DefaultAddresses())
+		anonymizer.SetLevel(opts.AnonymizeLevel)
 		anonymizeOverview(anonymizer, &overview)
 	}
 
@@ -976,6 +980,7 @@ func timeAgo(t time.Time) string {
 
 func anonymizePeerDetail(a *anonymize.Anonymizer, peer *PeerStateDetailOutput) {
 	peer.FQDN = a.AnonymizeDomain(peer.FQDN)
+	peer.PubKey = a.AnonymizeWGKey(peer.PubKey)
 	if localIP, port, err := net.SplitHostPort(peer.IceCandidateEndpoint.Local); err == nil {
 		peer.IceCandidateEndpoint.Local = fmt.Sprintf("%s:%s", a.AnonymizeIPString(localIP), port)
 	}
@@ -1007,6 +1012,7 @@ func anonymizeOverview(a *anonymize.Anonymizer, overview *OutputOverview) {
 	overview.SignalState.URL = a.AnonymizeURI(overview.SignalState.URL)
 	overview.SignalState.Error = a.AnonymizeString(overview.SignalState.Error)
 
+	overview.PubKey = a.AnonymizeWGKey(overview.PubKey)
 	overview.IP = a.AnonymizeIPString(overview.IP)
 	overview.IPv6 = a.AnonymizeIPString(overview.IPv6)
 	for i, detail := range overview.Relays.Details {