Authorize daemon IPC callers by their local identity

CONTRIBUTING.md

@@ -234,12 +234,22 @@ cd client/ui
 task dev
 ```
 
-Pass daemon flags after `--`:
+Pass daemon flags after `--`, pointing the UI at the socket the daemon serves:
 
 ```
-task dev -- --daemon-addr=tcp://127.0.0.1:41731
+task dev -- --daemon-addr=unix:///var/run/netbird.sock   # Linux, macOS
+task dev -- --daemon-addr=npipe://netbird                # Windows
 ```
 
+On Windows the daemon serves a named pipe (`npipe://netbird`). Which path that
+ends up being depends on what the daemon may create: as a service or elevated it
+serves `\\.\pipe\ProtectedPrefix\Administrators\netbird`, which no unprivileged
+process can take from it, and otherwise it falls back to `\\.\pipe\netbird`.
+Clients try both and check who owns the pipe before using the plain one. Avoid
+`tcp://127.0.0.1:41731`: loopback TCP carries no caller identity, so the daemon
+refuses the operations that require an administrator and you will not exercise
+those paths.
+
 Production build (frontend assets embedded into the binary, output in `client/ui/bin/`):
 
 ```

client/cmd/daemon_error.go

@@ -0,0 +1,66 @@
+package cmd
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+
+	"google.golang.org/genproto/googleapis/rpc/errdetails"
+	gstatus "google.golang.org/grpc/status"
+
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
+)
+
+// daemonCallError prepares a daemon error for display. A refusal the daemon
+// raised because the operation needs root/administrator is already guidance
+// written for the user, so it is surfaced on its own instead of buried under the
+// gRPC envelope and the name of the RPC that hit it. Anything else is wrapped
+// with context as usual.
+func daemonCallError(context string, err error) error {
+	if guidance, ok := privilegeGuidance(err); ok {
+		return errors.New(guidance)
+	}
+	return fmt.Errorf("%s: %w", context, err)
+}
+
+// privilegeGuidance renders the daemon's privilege refusal as a summary and the
+// command that performs the operation with the privileges it needs. It reports
+// false for any other error.
+func privilegeGuidance(err error) (string, bool) {
+	info, ok := privilegeErrorInfo(err)
+	if !ok {
+		return "", false
+	}
+
+	summary := info.GetMetadata()[ipcauth.ErrorMetaSummary]
+	command := info.GetMetadata()[ipcauth.ErrorMetaCommand]
+	if summary == "" {
+		// Detail without a summary: fall back to the status message, which
+		// carries the same text.
+		summary = strings.TrimSpace(gstatus.Convert(err).Message())
+	}
+	if command == "" {
+		return summary, true
+	}
+
+	return fmt.Sprintf("%s\n\n    %s\n", summary, command), true
+}
+
+// privilegeErrorInfo returns the daemon's privilege-refusal detail, if the error
+// carries one.
+func privilegeErrorInfo(err error) (*errdetails.ErrorInfo, bool) {
+	if err == nil {
+		return nil, false
+	}
+
+	for _, detail := range gstatus.Convert(err).Details() {
+		info, ok := detail.(*errdetails.ErrorInfo)
+		if !ok {
+			continue
+		}
+		if info.GetReason() == ipcauth.ErrorReasonPrivilegeRequired && info.GetDomain() == ipcauth.ErrorDomain {
+			return info, true
+		}
+	}
+	return nil, false
+}

client/cmd/logout.go

@@ -46,7 +46,7 @@ var logoutCmd = &cobra.Command{
 		}
 
 		if _, err := daemonClient.Logout(ctx, req); err != nil {
-			return fmt.Errorf("deregister: %v", err)
+			return daemonCallError("deregister", err)
 		}
 
 		cmd.Println("Deregistered successfully")

client/cmd/root.go

@@ -20,7 +20,6 @@ import (
 	"github.com/spf13/cobra"
 	"github.com/spf13/pflag"
 	"google.golang.org/grpc"
-	"google.golang.org/grpc/credentials/insecure"
 
 	daddr "github.com/netbirdio/netbird/client/internal/daemonaddr"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
@@ -91,6 +90,7 @@ var (
 			// Don't resolve for service commands — they create the socket, not connect to it.
 			if !isServiceCmd(cmd) {
 				daemonAddr = daddr.ResolveUnixDaemonAddr(daemonAddr)
+				daemonAddr = daddr.ResolveDaemonAddr(daemonAddr)
 			}
 			return nil
 		},
@@ -143,10 +143,10 @@ func init() {
 
 	defaultDaemonAddr := "unix:///var/run/netbird.sock"
 	if runtime.GOOS == "windows" {
-		defaultDaemonAddr = "tcp://127.0.0.1:41731"
+		defaultDaemonAddr = daddr.WindowsPipeAddr
 	}
 
-	rootCmd.PersistentFlags().StringVar(&daemonAddr, "daemon-addr", defaultDaemonAddr, "Daemon service address to serve CLI requests [unix|tcp]://[path|host:port]")
+	rootCmd.PersistentFlags().StringVar(&daemonAddr, "daemon-addr", defaultDaemonAddr, "Daemon service address to serve CLI requests [unix|tcp|npipe]://[path|host:port|name]")
 	rootCmd.PersistentFlags().StringVarP(&managementURL, "management-url", "m", "", fmt.Sprintf("Management Service URL [http|https]://[host]:[port] (default \"%s\")", profilemanager.DefaultManagementURL))
 	rootCmd.PersistentFlags().StringVar(&adminURL, "admin-url", "", fmt.Sprintf("Admin Panel URL [http|https]://[host]:[port] (default \"%s\")", profilemanager.DefaultAdminURL))
 	rootCmd.PersistentFlags().StringVarP(&logLevel, "log-level", "l", "info", "sets NetBird log level")
@@ -269,12 +269,10 @@ func DialClientGRPCServer(ctx context.Context, addr string) (*grpc.ClientConn, e
 	ctx, cancel := context.WithTimeout(ctx, time.Second*10)
 	defer cancel()
 
-	return grpc.DialContext(
-		ctx,
-		strings.TrimPrefix(addr, "tcp://"),
-		grpc.WithTransportCredentials(insecure.NewCredentials()),
-		grpc.WithBlock(),
-	)
+	target, opts := daddr.DialTarget(addr)
+	opts = append(opts, grpc.WithBlock())
+
+	return grpc.DialContext(ctx, target, opts...)
 }
 
 // WithBackOff execute function in backoff cycle.

client/cmd/service.go

@@ -33,10 +33,15 @@ var (
 )
 
 type program struct {
-	ctx              context.Context
-	cancel           context.CancelFunc
-	serv             *grpc.Server
-	jsonServ         *http.Server
+	ctx      context.Context
+	cancel   context.CancelFunc
+	serv     *grpc.Server
+	jsonServ *http.Server
+	// jsonClient is the gateway's own connection to the daemon. It is held so
+	// shutting the gateway down also closes it: nothing else references it once
+	// the handlers are registered, so its transport goroutines would otherwise
+	// outlive the server.
+	jsonClient       *grpc.ClientConn
 	jsonServMu       sync.Mutex
 	serverInstance   *server.Server
 	serverInstanceMu sync.Mutex

client/cmd/service_controller.go

@@ -5,6 +5,7 @@ package cmd
 import (
 	"context"
 	"fmt"
+	"runtime"
 	"time"
 
 	"github.com/kardianos/service"
@@ -13,6 +14,8 @@ import (
 	"github.com/spf13/cobra"
 	"google.golang.org/grpc"
 
+	"github.com/netbirdio/netbird/client/internal/daemonaddr"
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
 	"github.com/netbirdio/netbird/client/proto"
 	"github.com/netbirdio/netbird/client/server"
 	"github.com/netbirdio/netbird/client/system"
@@ -26,6 +29,31 @@ func validateJSONSocketFlags() error {
 	return nil
 }
 
+// daemonServerOptions installs the transport credentials that expose each
+// caller's kernel-authenticated identity to the handlers, which is what lets
+// the daemon require root/administrator for privileged operations.
+//
+// The handshake exchanges no bytes, so older CLI and UI binaries still
+// interoperate. Callers on a TCP socket carry no identity at all: the daemon
+// keeps serving them, and the privileged operations deny them, so a warning is
+// logged to make the loss of functionality visible.
+func daemonServerOptions(network string) []grpc.ServerOption {
+	if network == "tcp" {
+		log.Warnf("daemon is listening on TCP (%s): callers carry no verifiable identity over TCP, "+
+			"so privileged operations (SSH root login, SSH auth, enabling the SSH server, management URL changes, "+
+			"deregistration) will be denied. Use a unix socket, or npipe:// on Windows", daemonAddr)
+		return nil
+	}
+
+	creds := ipcauth.NewTransportCredentials()
+	if creds == nil {
+		log.Warnf("daemon IPC has no peer-identity primitive on %s: privileged operations will be denied", runtime.GOOS)
+		return nil
+	}
+
+	return []grpc.ServerOption{grpc.Creds(creds)}
+}
+
 func (p *program) Start(svc service.Service) error {
 	// Start should not block. Do the actual work async.
 	log.Info("starting NetBird service") //nolint
@@ -37,68 +65,106 @@ func (p *program) Start(svc service.Service) error {
 	// Collect static system and platform information
 	system.UpdateStaticInfoAsync()
 
-	// in any case, even if configuration does not exists we run daemon to serve CLI gRPC API.
-	p.serv = grpc.NewServer()
-
-	daemonListener, err := listenOnAddress(daemonAddr)
-	if err != nil {
-		return fmt.Errorf("listen daemon interface: %w", err)
+	// A daemon installed before named-pipe support has the loopback TCP address
+	// persisted. Move it to the named pipe so an upgraded daemon can identify
+	// its callers instead of silently serving an unauthenticated socket.
+	if migrated, ok := daemonaddr.MigrateLegacy(daemonAddr); ok {
+		log.Infof("daemon address %q predates named-pipe support, listening on %q so callers can be identified", daemonAddr, migrated)
+		daemonAddr = migrated
 	}
 
-	var jsonListener *socketListener
-	if enableJSONSocket {
-		jsonListener, err = listenOnAddress(jsonSocket)
-		if err != nil {
-			_ = daemonListener.Close()
-			return fmt.Errorf("listen daemon JSON interface: %w", err)
-		}
-	} else {
-		removeStaleUnixSocketForAddress(jsonSocket)
+	network, _, err := parseListenAddress(daemonAddr)
+	if err != nil {
+		return fmt.Errorf("parse daemon address: %w", err)
+	}
+
+	// in any case, even if configuration does not exists we run daemon to serve CLI gRPC API.
+	p.serv = grpc.NewServer(daemonServerOptions(network)...)
+
+	daemonListener, jsonListener, err := listenDaemonSockets()
+	if err != nil {
+		return err
 	}
 
 	go func() {
-		defer daemonListener.Close()
-		if jsonListener != nil {
-			defer jsonListener.Close()
-		}
-
-		if err := daemonListener.chmodUnixSocket("daemon"); err != nil {
-			log.Error(err)
-			return
-		}
-		if jsonListener != nil {
-			if err := jsonListener.chmodUnixSocket("daemon JSON"); err != nil {
-				log.Error(err)
-				return
-			}
-		}
-
-		serverInstance := server.New(p.ctx, util.FindFirstLogPath(logFiles), configPath, profilesDisabled, updateSettingsDisabled, captureEnabled, networksDisabled)
-		if err := serverInstance.Start(); err != nil {
-			log.Fatalf("failed to start daemon: %v", err)
-		}
-		proto.RegisterDaemonServiceServer(p.serv, serverInstance)
-
-		p.serverInstanceMu.Lock()
-		p.serverInstance = serverInstance
-		p.serverInstanceMu.Unlock()
-
-		if jsonListener != nil {
-			if err := p.startJSONGateway(jsonListener, daemonAddr); err != nil {
-				log.Fatalf("failed to start daemon JSON server: %v", err)
-			}
-		} else {
-			log.Debug("daemon JSON socket disabled")
-		}
-
-		log.Printf("started daemon server: %v", daemonListener.address)
-		if err := p.serv.Serve(daemonListener.Listener); err != nil {
-			log.Errorf("failed to serve daemon requests: %v", err)
+		// Fatal here rather than inside serve, so serve's deferred listener
+		// closes run before the process exits.
+		if err := p.serve(daemonListener, jsonListener); err != nil {
+			log.Fatalf("failed to %v", err)
 		}
 	}()
 	return nil
 }
 
+// listenDaemonSockets opens the daemon control socket and, when it is enabled, the
+// JSON gateway socket. The control socket is closed again if the second one fails,
+// so a failed start leaves nothing listening. The returned JSON listener is nil
+// when the socket is disabled.
+func listenDaemonSockets() (*socketListener, *socketListener, error) {
+	daemonListener, err := listenOnAddress(daemonAddr)
+	if err != nil {
+		return nil, nil, fmt.Errorf("listen daemon interface: %w", err)
+	}
+
+	if !enableJSONSocket {
+		removeStaleUnixSocketForAddress(jsonSocket)
+		return daemonListener, nil, nil
+	}
+
+	jsonListener, err := listenOnAddress(jsonSocket)
+	if err != nil {
+		if cerr := daemonListener.Close(); cerr != nil {
+			log.Debugf("close daemon listener: %v", cerr)
+		}
+		return nil, nil, fmt.Errorf("listen daemon JSON interface: %w", err)
+	}
+
+	return daemonListener, jsonListener, nil
+}
+
+// serve brings up the daemon server on an already-open control socket and blocks
+// until it stops. jsonListener is nil when the JSON socket is disabled. A returned
+// error means the daemon cannot run at all and the caller is expected to exit; the
+// failures it recovers from on its own are logged here.
+func (p *program) serve(daemonListener, jsonListener *socketListener) error {
+	defer daemonListener.Close()
+	if jsonListener != nil {
+		defer jsonListener.Close()
+	}
+
+	// chmodUnixSocket is a no-op for a nil listener and for a non-unix one.
+	if err := daemonListener.chmodUnixSocket("daemon"); err != nil {
+		log.Error(err)
+		return nil
+	}
+	if err := jsonListener.chmodUnixSocket("daemon JSON"); err != nil {
+		log.Error(err)
+		return nil
+	}
+
+	serverInstance := server.New(p.ctx, util.FindFirstLogPath(logFiles), configPath, profilesDisabled, updateSettingsDisabled, captureEnabled, networksDisabled)
+	if err := serverInstance.Start(); err != nil {
+		return fmt.Errorf("start daemon: %w", err)
+	}
+	proto.RegisterDaemonServiceServer(p.serv, serverInstance)
+
+	p.serverInstanceMu.Lock()
+	p.serverInstance = serverInstance
+	p.serverInstanceMu.Unlock()
+
+	if jsonListener == nil {
+		log.Debug("daemon JSON socket disabled")
+	} else if err := p.startJSONGateway(jsonListener, daemonAddr); err != nil {
+		return fmt.Errorf("start daemon JSON server: %w", err)
+	}
+
+	log.Printf("started daemon server: %v", daemonListener.address)
+	if err := p.serv.Serve(daemonListener.Listener); err != nil {
+		log.Errorf("failed to serve daemon requests: %v", err)
+	}
+	return nil
+}
+
 func (p *program) Stop(srv service.Service) error {
 	p.serverInstanceMu.Lock()
 	if p.serverInstance != nil {
@@ -113,8 +179,13 @@ func (p *program) Stop(srv service.Service) error {
 	p.cancel()
 
 	p.jsonServMu.Lock()
-	jsonServ := p.jsonServ
+	jsonServ, jsonClient := p.jsonServ, p.jsonClient
 	p.jsonServMu.Unlock()
+	if jsonClient != nil {
+		if err := jsonClient.Close(); err != nil {
+			log.Debugf("close daemon JSON gateway client: %v", err)
+		}
+	}
 	if jsonServ != nil {
 		shutdownCtx, shutdownCancel := context.WithTimeout(context.Background(), 2*time.Second)
 		if err := jsonServ.Shutdown(shutdownCtx); err != nil {

client/cmd/service_json_gateway.go

@@ -5,27 +5,123 @@ package cmd
 import (
 	"context"
 	"errors"
+	"fmt"
 	"net"
 	"net/http"
-	"strings"
+	"sync"
 	"time"
 
 	"github.com/grpc-ecosystem/grpc-gateway/v2/runtime"
 	log "github.com/sirupsen/logrus"
 	"google.golang.org/grpc"
-	"google.golang.org/grpc/credentials/insecure"
 
+	"github.com/netbirdio/netbird/client/internal/daemonaddr"
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
 	"github.com/netbirdio/netbird/client/proto"
 )
 
-func grpcGatewayEndpoint(addr string) string {
-	return strings.TrimPrefix(addr, "tcp://")
+// jsonPeerIdentity is the context key under which the connecting HTTP client's
+// identity is stashed for the lifetime of its connection.
+type jsonPeerIdentity struct{}
+
+// jsonPeerIdentityValue pairs the identity with whether it could be read at
+// all, so an unreadable identity is forwarded as "unknown" rather than omitted.
+type jsonPeerIdentityValue struct {
+	id    ipcauth.Identity
+	known bool
+}
+
+// jsonConnContext reads the identity of the client connecting to the JSON
+// socket and stashes it on the connection's context. The gateway re-dials the
+// daemon in-process, so the daemon would otherwise see every JSON request as
+// coming from the daemon itself.
+func jsonConnContext(ctx context.Context, c net.Conn) context.Context {
+	value := jsonPeerIdentityValue{}
+	id, err := ipcauth.ConnIdentity(c)
+	if err != nil {
+		log.Warnf("json gateway: cannot read HTTP client identity, privileged operations will be denied for this connection: %v", err)
+	} else {
+		value.id = id
+		value.known = true
+	}
+	return context.WithValue(ctx, jsonPeerIdentity{}, value)
+}
+
+// forwardIdentity stamps the HTTP client's identity onto every call the gateway
+// makes to the daemon.
+//
+// It is an interceptor on the gateway's client connection rather than a
+// runtime.WithMetadata annotator because grpc-gateway skips annotators when no
+// request header maps to metadata, which an HTTP/1.0 request with no Host header
+// over a unix socket achieves. The daemon would then receive no marker, see its own
+// identity as the transport peer, and authorize the request as the daemon itself.
+// An interceptor runs for every RPC whatever the request looked like.
+func forwardIdentity(ctx context.Context) context.Context {
+	value, ok := ctx.Value(jsonPeerIdentity{}).(jsonPeerIdentityValue)
+	if !ok {
+		// No ConnContext ran for this request, so forward an unknown identity:
+		// the daemon must not mistake its own identity for the client's.
+		return ipcauth.WithForwardedIdentity(ctx, ipcauth.Identity{}, false)
+	}
+	return ipcauth.WithForwardedIdentity(ctx, value.id, value.known)
+}
+
+func forwardIdentityUnary(ctx context.Context, method string, req, reply any, cc *grpc.ClientConn, invoker grpc.UnaryInvoker, opts ...grpc.CallOption) error {
+	return invoker(forwardIdentity(ctx), method, req, reply, cc, opts...)
+}
+
+func forwardIdentityStream(ctx context.Context, desc *grpc.StreamDesc, cc *grpc.ClientConn, method string, streamer grpc.Streamer, opts ...grpc.CallOption) (grpc.ClientStream, error) {
+	return streamer(forwardIdentity(ctx), desc, cc, method, opts...)
+}
+
+// reservedHeaderWarning limits the dropped-header warning to the first occurrence.
+var reservedHeaderWarning sync.Once
+
+// jsonIncomingHeaderMatcher keeps an HTTP client from supplying the metadata the
+// gateway uses to forward its identity. grpc-gateway turns "Grpc-Metadata-<key>"
+// headers into gRPC metadata and joins them ahead of what its annotators add, so
+// without this filter a JSON client could send its own x-netbird-fwd-uid and the
+// daemon would authorize that instead of the client's real identity.
+func jsonIncomingHeaderMatcher(key string) (string, bool) {
+	mapped, ok := runtime.DefaultHeaderMatcher(key)
+	if !ok {
+		return "", false
+	}
+	if ipcauth.IsReservedForwardKey(mapped) {
+		// Warn once: any client can send these on every request, so warning each
+		// time hands it a way to fill the log. The rest are debug-level.
+		reservedHeaderWarning.Do(func() {
+			log.Warnf("json gateway: dropping reserved header %q from a request: only the gateway may set the caller's identity", key)
+		})
+		log.Debugf("json gateway: dropping reserved header %q", key)
+		return "", false
+	}
+	return mapped, true
 }
 
 func (p *program) startJSONGateway(jsonListener *socketListener, daemonEndpoint string) error {
-	mux := runtime.NewServeMux()
-	opts := []grpc.DialOption{grpc.WithTransportCredentials(insecure.NewCredentials())}
-	if err := proto.RegisterDaemonServiceHandlerFromEndpoint(p.ctx, mux, grpcGatewayEndpoint(daemonEndpoint), opts); err != nil {
+	if jsonListener.network == "tcp" {
+		log.Warnf("daemon JSON socket is listening on TCP (%s): callers carry no verifiable identity over TCP, "+
+			"so privileged operations will be denied for JSON clients", jsonListener.address)
+	}
+
+	mux := runtime.NewServeMux(runtime.WithIncomingHeaderMatcher(jsonIncomingHeaderMatcher))
+
+	// grpc.NewClient does not connect until the first request, so registering
+	// the handler here cannot block daemon startup.
+	target, opts := daemonaddr.DialTarget(daemonEndpoint)
+	opts = append(opts,
+		grpc.WithChainUnaryInterceptor(forwardIdentityUnary),
+		grpc.WithChainStreamInterceptor(forwardIdentityStream),
+	)
+	conn, err := grpc.NewClient(target, opts...)
+	if err != nil {
+		return fmt.Errorf("create daemon client for JSON gateway: %w", err)
+	}
+	if err := proto.RegisterDaemonServiceHandler(p.ctx, mux, conn); err != nil {
+		if cerr := conn.Close(); cerr != nil {
+			log.Debugf("close daemon client after failed JSON gateway registration: %v", cerr)
+		}
 		return err
 	}
 
@@ -35,10 +131,12 @@ func (p *program) startJSONGateway(jsonListener *socketListener, daemonEndpoint
 		BaseContext: func(net.Listener) context.Context {
 			return p.ctx
 		},
+		ConnContext: jsonConnContext,
 	}
 
 	p.jsonServMu.Lock()
 	p.jsonServ = jsonServer
+	p.jsonClient = conn
 	p.jsonServMu.Unlock()
 
 	go func() {

client/cmd/service_json_gateway_test.go

@@ -0,0 +1,261 @@
+//go:build !windows && !ios && !android
+
+package cmd
+
+import (
+	"context"
+	"net"
+	"net/http"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"github.com/grpc-ecosystem/grpc-gateway/v2/runtime"
+	"google.golang.org/grpc/credentials"
+	"google.golang.org/grpc/metadata"
+	"google.golang.org/grpc/peer"
+
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
+)
+
+// The JSON gateway runs inside the daemon and re-dials it locally, so every JSON
+// request reaches a handler with the daemon's own identity as the transport peer.
+// The gateway therefore forwards its HTTP client's identity as metadata, and the
+// daemon authorizes that instead of itself. These tests drive the real wiring
+// (jsonConnContext, forwardIdentity, jsonIncomingHeaderMatcher) and check the
+// identity a handler would end up authorizing.
+
+// daemonSideCtx is what a handler sees for a gateway-relayed call. The transport
+// peer must be this process's own identity: the gateway is the daemon, so the two
+// cannot differ, and hardcoding root here instead would describe a state that
+// never occurs.
+func daemonSideCtx(t *testing.T, md metadata.MD) context.Context {
+	t.Helper()
+	self, err := ipcauth.CurrentProcessIdentity()
+	if err != nil {
+		t.Skipf("cannot read this process's identity: %v", err)
+	}
+	ctx := peer.NewContext(context.Background(), &peer.Peer{
+		AuthInfo: ipcauth.AuthInfo{
+			CommonAuthInfo: credentials.CommonAuthInfo{SecurityLevel: credentials.NoSecurity},
+			Identity:       self,
+		},
+	})
+	return metadata.NewIncomingContext(ctx, md)
+}
+
+// gatewayMetadata reproduces what the daemon receives for a JSON request: the
+// mux annotates the context from the request's headers, then the interceptor on the
+// gateway's client connection stamps the caller's identity. The order matters,
+// since the interceptor must win over anything a header put there.
+func gatewayMetadata(t *testing.T, req *http.Request, ctx context.Context) metadata.MD {
+	t.Helper()
+	mux := runtime.NewServeMux(runtime.WithIncomingHeaderMatcher(jsonIncomingHeaderMatcher))
+	annotated, err := runtime.AnnotateContext(ctx, mux, req,
+		"/daemon.DaemonService/SetConfig",
+		runtime.WithHTTPPathPattern("/daemon.DaemonService/SetConfig"))
+	if err != nil {
+		t.Fatalf("annotate: %v", err)
+	}
+
+	md, ok := metadata.FromOutgoingContext(forwardIdentity(annotated))
+	if !ok {
+		t.Fatal("the interceptor produced no metadata")
+	}
+	return md
+}
+
+// clientCtx is the connection context jsonConnContext would have produced for an
+// HTTP client whose identity the gateway could read.
+func clientCtx(id ipcauth.Identity, known bool) context.Context {
+	return context.WithValue(context.Background(), jsonPeerIdentity{},
+		jsonPeerIdentityValue{id: id, known: known})
+}
+
+// An HTTP client must not be able to name its own identity. grpc-gateway turns
+// Grpc-Metadata-<key> headers into gRPC metadata, so without the header filter and
+// the interceptor overwriting the reserved keys, this request would authorize as
+// uid 0.
+func TestJSONGateway_ForgedIdentityHeaderIsDropped(t *testing.T) {
+	req, err := http.NewRequest(http.MethodPost, "http://localhost/daemon.DaemonService/SetConfig", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	req.Header.Set("Grpc-Metadata-X-Netbird-Fwd-Uid", "0")
+	req.Header.Set("Grpc-Metadata-X-Netbird-Fwd-Gid", "0")
+	req.Header.Set("Grpc-Metadata-X-Netbird-Fwd", "1")
+	req.Header.Set("Grpc-Metadata-X-Netbird-Fwd-Sid", "S-1-5-18")
+
+	caller := ipcauth.Identity{UID: 31000, GID: 31000}
+	md := gatewayMetadata(t, req, clientCtx(caller, true))
+
+	id, ok := ipcauth.CallerIdentity(daemonSideCtx(t, md))
+	if !ok {
+		t.Fatal("the forwarded identity should be usable")
+	}
+	if id.IsPrivileged() {
+		t.Errorf("forged header was believed: authorized as %v", id)
+	}
+	if id.UID != caller.UID {
+		t.Errorf("authorized as uid %d, want the real client %d", id.UID, caller.UID)
+	}
+}
+
+// A request with no headers at all (HTTP/1.0 needs no Host, and a unix socket
+// yields no host:port) makes grpc-gateway produce no metadata whatsoever and skip
+// its annotators: "if len(pairs) == 0 { return ctx, nil, nil }" in
+// runtime/context.go. That is why the identity is stamped by an interceptor
+// instead. This is the case that previously reached the gate as the daemon itself.
+func TestJSONGateway_HeaderlessRequestIsStillMarkedForwarded(t *testing.T) {
+	req, err := http.NewRequest(http.MethodPost, "http://localhost/daemon.DaemonService/SetConfig", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	req.Header = http.Header{}
+	req.Host = ""
+
+	caller := ipcauth.Identity{UID: 31000, GID: 31000}
+	ctx := clientCtx(caller, true)
+
+	// Pin the skip path itself: if grpc-gateway ever produced a pair here, this
+	// test would still pass below while no longer covering what it was written for.
+	mux := runtime.NewServeMux(runtime.WithIncomingHeaderMatcher(jsonIncomingHeaderMatcher))
+	annotated, err := runtime.AnnotateContext(ctx, mux, req,
+		"/daemon.DaemonService/SetConfig",
+		runtime.WithHTTPPathPattern("/daemon.DaemonService/SetConfig"))
+	if err != nil {
+		t.Fatalf("annotate: %v", err)
+	}
+	if md, ok := metadata.FromOutgoingContext(annotated); ok {
+		t.Fatalf("grpc-gateway produced metadata %v for a headerless request; "+
+			"this test no longer covers the annotator-skip path", md)
+	}
+
+	md := gatewayMetadata(t, req, ctx)
+
+	id, ok := ipcauth.CallerIdentity(daemonSideCtx(t, md))
+	if !ok {
+		t.Fatal("the forwarded identity should be usable")
+	}
+	if id.UID != caller.UID || id.IsPrivileged() {
+		t.Errorf("authorized as %v, want the real client uid %d", id, caller.UID)
+	}
+}
+
+// When the gateway cannot read its client's identity (a TCP JSON socket, say) it
+// forwards the marker alone. The daemon must then report "unidentified" so the
+// privileged operations refuse, rather than falling back to the gateway's own
+// identity.
+func TestJSONGateway_UnreadableClientIdentityIsUnidentified(t *testing.T) {
+	req, err := http.NewRequest(http.MethodPost, "http://localhost/daemon.DaemonService/SetConfig", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	md := gatewayMetadata(t, req, clientCtx(ipcauth.Identity{}, false))
+
+	if id, ok := ipcauth.CallerIdentity(daemonSideCtx(t, md)); ok {
+		t.Errorf("a request with no client identity was authorized as %v", id)
+	}
+}
+
+// A request that never passed through jsonConnContext (no stashed identity) must
+// also come out unidentified rather than as the daemon.
+func TestJSONGateway_MissingConnContextIsUnidentified(t *testing.T) {
+	req, err := http.NewRequest(http.MethodPost, "http://localhost/daemon.DaemonService/SetConfig", nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	md := gatewayMetadata(t, req, context.Background())
+
+	if id, ok := ipcauth.CallerIdentity(daemonSideCtx(t, md)); ok {
+		t.Errorf("a request with no connection context was authorized as %v", id)
+	}
+}
+
+// End to end over a real unix socket: the gateway reads the connecting client's
+// identity from the socket itself, so a client cannot present anything else.
+func TestJSONGateway_IdentityComesFromTheSocket(t *testing.T) {
+	mux := runtime.NewServeMux(runtime.WithIncomingHeaderMatcher(jsonIncomingHeaderMatcher))
+
+	type observed struct {
+		md metadata.MD
+	}
+	seen := make(chan observed, 1)
+
+	srv := &http.Server{
+		Handler: http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+			ctx, err := runtime.AnnotateContext(r.Context(), mux, r,
+				"/daemon.DaemonService/SetConfig",
+				runtime.WithHTTPPathPattern("/daemon.DaemonService/SetConfig"))
+			if err != nil {
+				t.Errorf("annotate: %v", err)
+				return
+			}
+			md, _ := metadata.FromOutgoingContext(forwardIdentity(ctx))
+			seen <- observed{md: md}
+			w.WriteHeader(http.StatusOK)
+		}),
+		ReadHeaderTimeout: 5 * time.Second,
+		ConnContext:       jsonConnContext,
+	}
+
+	sock := filepath.Join(t.TempDir(), "http.sock")
+	ln, err := net.Listen("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() {
+		if err := srv.Close(); err != nil {
+			t.Logf("close server: %v", err)
+		}
+	})
+	go func() {
+		if err := srv.Serve(ln); err != nil && err != http.ErrServerClosed {
+			t.Logf("serve: %v", err)
+		}
+	}()
+
+	conn, err := net.Dial("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() {
+		if err := conn.Close(); err != nil {
+			t.Logf("close conn: %v", err)
+		}
+	})
+
+	// Forge the identity headers on the wire as well.
+	request := "POST /daemon.DaemonService/SetConfig HTTP/1.1\r\n" +
+		"Host: localhost\r\n" +
+		"Grpc-Metadata-X-Netbird-Fwd: 1\r\n" +
+		"Grpc-Metadata-X-Netbird-Fwd-Uid: 0\r\n" +
+		"Content-Length: 0\r\n\r\n"
+	if _, err := conn.Write([]byte(request)); err != nil {
+		t.Fatal(err)
+	}
+
+	select {
+	case got := <-seen:
+		self, err := ipcauth.CurrentProcessIdentity()
+		if err != nil {
+			t.Skipf("cannot read this process's identity: %v", err)
+		}
+		// The socket peer is this test process, so that is the identity the
+		// gateway must forward, not the uid 0 the request asked for.
+		if uids := got.md.Get("x-netbird-fwd-uid"); len(uids) != 1 {
+			t.Fatalf("x-netbird-fwd-uid = %v, want exactly the gateway's own value", uids)
+		}
+		id, ok := ipcauth.CallerIdentity(daemonSideCtx(t, got.md))
+		if !ok {
+			t.Fatal("the forwarded identity should be usable")
+		}
+		if id.UID != self.UID {
+			t.Errorf("authorized as uid %d, want the socket peer %d", id.UID, self.UID)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("the gateway never handled the request")
+	}
+}

client/cmd/service_params.go

@@ -13,6 +13,7 @@ import (
 	"github.com/spf13/cobra"
 
 	"github.com/netbirdio/netbird/client/configs"
+	"github.com/netbirdio/netbird/client/internal/daemonaddr"
 	"github.com/netbirdio/netbird/util"
 )
 
@@ -125,6 +126,13 @@ func applyServiceParams(cmd *cobra.Command, params *serviceParams) {
 
 	if !rootCmd.PersistentFlags().Changed("daemon-addr") && params.DaemonAddr != "" {
 		daemonAddr = params.DaemonAddr
+		// An install that predates named-pipe support has the loopback TCP
+		// address saved. Callers carry no identity over TCP, so move it to the
+		// pipe instead of restoring a socket the daemon cannot authorize on.
+		if migrated, ok := daemonaddr.MigrateLegacy(daemonAddr); ok {
+			cmd.Printf("Moving the saved daemon address from %s to %s so the daemon can identify its callers\n", daemonAddr, migrated)
+			daemonAddr = migrated
+		}
 	}
 
 	if !serviceCmd.PersistentFlags().Changed("json-socket") && params.JSONSocket != "" {

client/cmd/service_pipe_other.go

@@ -0,0 +1,14 @@
+//go:build !windows
+
+package cmd
+
+import (
+	"fmt"
+	"net"
+)
+
+// listenNamedPipe is Windows-only: no other platform serves the daemon on a
+// named pipe.
+func listenNamedPipe(string) (net.Listener, string, error) {
+	return nil, "", fmt.Errorf("named pipes are only supported on Windows")
+}

client/cmd/service_pipe_windows.go

@@ -0,0 +1,41 @@
+//go:build windows
+
+package cmd
+
+import (
+	"errors"
+	"fmt"
+	"net"
+
+	"github.com/Microsoft/go-winio"
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/daemonaddr"
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
+)
+
+// listenNamedPipe creates the daemon control pipe and reports the path it ended
+// up on. The security descriptor lets any local caller connect, as a Unix socket
+// at 0666 does, and the privileged operations are authorized separately from the
+// caller's token.
+//
+// The protected name comes first so that an unprivileged process cannot take the
+// name before the service does. Creating it requires being an administrator or
+// LocalSystem, so a daemon an ordinary user runs themselves, as in netstack mode,
+// falls back to the plain name; clients try both and check who serves them.
+func listenNamedPipe(name string) (net.Listener, string, error) {
+	var errs []error
+	for _, path := range daemonaddr.PipePaths(name) {
+		listener, err := winio.ListenPipe(path, &winio.PipeConfig{
+			SecurityDescriptor: ipcauth.DefaultPipeSDDL(),
+		})
+		if err != nil {
+			log.Debugf("not serving the daemon on %s: %v", path, err)
+			errs = append(errs, fmt.Errorf("%s: %w", path, err))
+			continue
+		}
+		return listener, path, nil
+	}
+
+	return nil, "", errors.Join(errs...)
+}

client/cmd/service_socket.go

@@ -26,6 +26,14 @@ func listenOnAddress(addr string) (*socketListener, error) {
 		return nil, err
 	}
 
+	if network == "npipe" {
+		listener, path, err := listenNamedPipe(address)
+		if err != nil {
+			return nil, err
+		}
+		return &socketListener{Listener: listener, network: network, address: path}, nil
+	}
+
 	if network == "unix" {
 		removeStaleUnixSocket(address)
 	}
@@ -41,11 +49,11 @@ func listenOnAddress(addr string) (*socketListener, error) {
 func parseListenAddress(addr string) (string, string, error) {
 	network, address, ok := strings.Cut(addr, "://")
 	if !ok || network == "" || address == "" {
-		return "", "", fmt.Errorf("address must be in [unix|tcp]://[path|host:port] format: %q", addr)
+		return "", "", fmt.Errorf("address must be in [unix|tcp|npipe]://[path|host:port|name] format: %q", addr)
 	}
 
 	switch network {
-	case "unix", "tcp":
+	case "unix", "tcp", "npipe":
 		return network, address, nil
 	default:
 		return "", "", fmt.Errorf("unsupported daemon address protocol: %v", network)

client/cmd/up.go

@@ -325,7 +325,7 @@ func runInDaemonMode(ctx context.Context, cmd *cobra.Command, pm *profilemanager
 		if st, ok := gstatus.FromError(err); ok && st.Code() == codes.Unavailable {
 			log.Warnf("setConfig method is not available in the daemon: %s", st.Message())
 		} else {
-			return fmt.Errorf("call service setConfig method: %v", err)
+			return daemonCallError("call service setConfig method", err)
 		}
 	}
 
@@ -379,7 +379,7 @@ func doDaemonUp(ctx context.Context, cmd *cobra.Command, client proto.DaemonServ
 	}
 
 	if loginErr != nil {
-		return fmt.Errorf("login failed: %v", loginErr)
+		return daemonCallError("login failed", loginErr)
 	}
 
 	if loginResp.NeedsSSOLogin {
@@ -392,7 +392,7 @@ func doDaemonUp(ctx context.Context, cmd *cobra.Command, client proto.DaemonServ
 		ProfileName: &profileID,
 		Username:    &username,
 	}); err != nil {
-		return fmt.Errorf("call service up method: %v", err)
+		return daemonCallError("call service up method", err)
 	}
 
 	return nil

client/internal/daemonaddr/owner.go

@@ -0,0 +1,15 @@
+package daemonaddr
+
+// DaemonRunsAsSelf reports whether the daemon listening at addr runs as this very
+// user. That is what makes an unprivileged daemon authorize this process for the
+// changes it otherwise restricts to root or an administrator, so a client can tell
+// up front whether those controls are usable instead of letting a save fail.
+//
+// It is answered from the ownership of the socket or pipe the daemon created, so it
+// costs no round trip and needs no cooperation from the daemon. Ownership that
+// cannot be read is reported as false, including for a TCP address, so a caller
+// reading this as "the daemon would allow it" fails closed. The daemon remains the
+// only thing that authorizes anything: this only decides what a client offers.
+func DaemonRunsAsSelf(addr string) bool {
+	return daemonRunsAsSelf(addr)
+}

client/internal/daemonaddr/owner_unix.go

@@ -0,0 +1,40 @@
+//go:build !windows
+
+package daemonaddr
+
+import (
+	"os"
+	"strings"
+	"syscall"
+
+	log "github.com/sirupsen/logrus"
+)
+
+// daemonRunsAsSelf compares the owner of the daemon's Unix socket with this
+// process's uid. Root is not treated specially here: a root caller is privileged
+// on its own merits, and a root-owned socket says nothing about the caller.
+func daemonRunsAsSelf(addr string) bool {
+	path, ok := strings.CutPrefix(addr, "unix://")
+	if !ok {
+		return false
+	}
+
+	info, err := os.Stat(path)
+	if err != nil {
+		log.Debugf("stat daemon socket %s: %v", path, err)
+		return false
+	}
+
+	// Only a socket says anything about a daemon. A directory or a leftover
+	// regular file at that path is not one, and reading it as "the daemon runs as
+	// us" would offer controls the daemon then refuses.
+	if info.Mode()&os.ModeSocket == 0 {
+		return false
+	}
+
+	stat, ok := info.Sys().(*syscall.Stat_t)
+	if !ok {
+		return false
+	}
+	return stat.Uid == uint32(os.Getuid())
+}

client/internal/daemonaddr/owner_unix_test.go

@@ -0,0 +1,62 @@
+//go:build !windows
+
+package daemonaddr
+
+import (
+	"net"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// A socket this user created means the daemon runs as this user, which is the
+// rootless case where the daemon delegates its authority to its own identity.
+func TestDaemonRunsAsSelf_OwnSocket(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "netbird.sock")
+	ln, err := net.Listen("unix", path)
+	if err != nil {
+		t.Fatalf("listen: %v", err)
+	}
+	t.Cleanup(func() {
+		if err := ln.Close(); err != nil {
+			t.Logf("close listener: %v", err)
+		}
+	})
+
+	if !DaemonRunsAsSelf("unix://" + path) {
+		t.Error("a socket owned by this user must count as the daemon running as us")
+	}
+}
+
+// Everything that is not a readable socket of ours has to answer false, because
+// the caller reads a true as "the daemon would authorize me".
+func TestDaemonRunsAsSelf_FailsClosed(t *testing.T) {
+	dir := t.TempDir()
+
+	// A socket owned by another user, which is what a root-run daemon looks like
+	// to an unprivileged client. Only assertable when we are not root ourselves.
+	rootOwned := "unix:///var/run/netbird.sock"
+	if _, err := os.Stat("/var/run/netbird.sock"); err == nil && os.Getuid() != 0 {
+		if DaemonRunsAsSelf(rootOwned) {
+			t.Error("a socket owned by another user must not count as ours")
+		}
+	}
+
+	for name, addr := range map[string]string{
+		"missing socket":  "unix://" + filepath.Join(dir, "absent.sock"),
+		"tcp address":     "tcp://127.0.0.1:41731",
+		"named pipe":      "npipe://netbird",
+		"empty":           "",
+		"no scheme":       filepath.Join(dir, "absent.sock"),
+		"directory":       "unix://" + dir,
+		"unknown scheme":  "http://localhost:8080",
+		"scheme only":     "unix://",
+		"relative socket": "unix://netbird.sock",
+	} {
+		t.Run(name, func(t *testing.T) {
+			if DaemonRunsAsSelf(addr) {
+				t.Errorf("%q must not count as a daemon running as us", addr)
+			}
+		})
+	}
+}

client/internal/daemonaddr/owner_windows.go

@@ -0,0 +1,42 @@
+//go:build windows
+
+package daemonaddr
+
+import (
+	"context"
+	"strings"
+
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
+)
+
+// daemonRunsAsSelf reads the owner of the daemon's pipe. A daemon running as the
+// service account owns its pipe as LocalSystem, and an elevated one as
+// BUILTIN\Administrators, so only a daemon the user started themselves matches.
+func daemonRunsAsSelf(addr string) bool {
+	name, ok := strings.CutPrefix(addr, pipeScheme)
+	if !ok {
+		return false
+	}
+
+	for _, path := range PipePaths(name) {
+		// Bounded: this runs on the UI's path for deciding which controls to
+		// offer, so a pipe that does not answer promptly must not stall it. A
+		// timeout leaves the caller unprivileged, which only disables controls.
+		ctx, cancel := context.WithTimeout(context.Background(), probeTimeout)
+		conn, err := dialPipe(ctx, path)
+		cancel()
+		if err != nil {
+			continue
+		}
+
+		owned := ipcauth.PipeOwnedBySelf(conn)
+		if cerr := conn.Close(); cerr != nil {
+			log.Debugf("close daemon pipe %s after ownership check: %v", path, cerr)
+		}
+		return owned
+	}
+
+	return false
+}

client/internal/daemonaddr/pipe.go

@@ -0,0 +1,103 @@
+package daemonaddr
+
+import (
+	"context"
+	"net"
+	"runtime"
+	"strings"
+
+	"google.golang.org/grpc"
+	"google.golang.org/grpc/credentials/insecure"
+)
+
+const (
+	// WindowsPipeAddr is the default daemon address on Windows. A named pipe
+	// carries the connecting process's token, which loopback TCP does not, so
+	// it is the only Windows transport on which the daemon can tell who is
+	// calling it.
+	WindowsPipeAddr = "npipe://netbird"
+
+	// legacyWindowsAddr is the loopback-TCP address the Windows daemon used
+	// before named-pipe support.
+	legacyWindowsAddr = "tcp://127.0.0.1:41731"
+
+	pipeScheme = "npipe://"
+
+	// protectedPrefix is the NPFS namespace in which only LocalSystem and
+	// members of BUILTIN\Administrators may create a pipe. A daemon running as
+	// the service account creates its pipe there so that an unprivileged process
+	// cannot pre-create the name, which would keep the daemon from starting and
+	// leave callers talking to the squatter. Opening such a pipe needs no
+	// privilege, so unprivileged clients still reach the daemon.
+	protectedPrefix = `ProtectedPrefix\Administrators\`
+)
+
+// DialTarget returns the gRPC dial target and transport options for a daemon
+// address. The npipe scheme needs a context dialer because gRPC has no
+// named-pipe resolver; unix and tcp are handled by gRPC itself.
+func DialTarget(addr string) (string, []grpc.DialOption) {
+	opts := []grpc.DialOption{grpc.WithTransportCredentials(insecure.NewCredentials())}
+
+	if name, ok := strings.CutPrefix(addr, pipeScheme); ok {
+		paths := PipePaths(name)
+		opts = append(opts, grpc.WithContextDialer(func(ctx context.Context, _ string) (net.Conn, error) {
+			return dialPipePaths(ctx, paths)
+		}))
+		return "passthrough:///netbird-daemon-pipe", opts
+	}
+
+	return strings.TrimPrefix(addr, "tcp://"), opts
+}
+
+// PipePath maps an npipe address name ("netbird", from "npipe://netbird") to a
+// Windows named-pipe path (\\.\pipe\netbird). A fully qualified path is left as
+// is.
+func PipePath(name string) string {
+	if strings.HasPrefix(name, `\\`) {
+		return name
+	}
+	return `\\.\pipe\` + name
+}
+
+// PipePaths returns the paths a daemon control pipe may live at for an npipe
+// address name, in the order both sides must try them: the protected name first,
+// then the plain one.
+//
+// The daemon serves the first it can create, which is the protected name when it
+// runs as the service account and the plain one when it runs as an ordinary user,
+// as it does in netstack mode. Clients therefore have to try both, and because a
+// client cannot tell from the name alone who created the pipe, the plain name is
+// only usable once the server's identity has been checked: see
+// verifyPipeServer.
+//
+// A fully qualified path is what the operator asked for and is used as is.
+func PipePaths(name string) []string {
+	if strings.HasPrefix(name, `\\`) {
+		return []string{name}
+	}
+	return []string{PipePath(protectedPrefix + name), PipePath(name)}
+}
+
+// IsProtectedPipePath reports whether a pipe path is in the namespace only an
+// administrator or LocalSystem can create in, which is what lets a client trust
+// such a pipe from its name alone.
+func IsProtectedPipePath(path string) bool {
+	return strings.HasPrefix(path, `\\.\pipe\`+protectedPrefix)
+}
+
+// MigrateLegacy upgrades the pre-named-pipe Windows daemon address to the named
+// pipe, reporting whether it rewrote the address. Existing installs persist the
+// daemon address, so without this an upgraded daemon would keep listening on
+// loopback TCP, where callers carry no identity and privileged operations would
+// have to be refused for everyone. Only the exact legacy default is rewritten:
+// a deliberately chosen custom address is left alone.
+func MigrateLegacy(addr string) (string, bool) {
+	return migrateLegacyForOS(runtime.GOOS, addr)
+}
+
+func migrateLegacyForOS(goos, addr string) (string, bool) {
+	if goos == "windows" && addr == legacyWindowsAddr {
+		return WindowsPipeAddr, true
+	}
+	return addr, false
+}

client/internal/daemonaddr/pipe_other.go

@@ -0,0 +1,15 @@
+//go:build !windows
+
+package daemonaddr
+
+import (
+	"context"
+	"fmt"
+	"net"
+)
+
+// dialPipePaths is Windows-only: no other platform serves the daemon on a named
+// pipe.
+func dialPipePaths(context.Context, []string) (net.Conn, error) {
+	return nil, fmt.Errorf("named pipes are only supported on Windows")
+}

client/internal/daemonaddr/pipe_test.go

@@ -0,0 +1,30 @@
+package daemonaddr
+
+import (
+	"slices"
+	"testing"
+)
+
+// The protected name must be tried before the plain one on both sides: it is the
+// one an unprivileged process cannot create, so preferring it is what keeps a
+// squatter from owning the name the service daemon would otherwise use.
+func TestPipePaths_PrefersTheProtectedName(t *testing.T) {
+	got := PipePaths("netbird")
+	want := []string{
+		`\\.\pipe\ProtectedPrefix\Administrators\netbird`,
+		`\\.\pipe\netbird`,
+	}
+	if !slices.Equal(got, want) {
+		t.Errorf("PipePaths = %q, want %q", got, want)
+	}
+}
+
+// An operator who passes a full path chose exactly one pipe, so neither side may
+// look anywhere else.
+func TestPipePaths_QualifiedPathIsUsedAsIs(t *testing.T) {
+	path := `\\.\pipe\custom-netbird`
+	got := PipePaths(path)
+	if !slices.Equal(got, []string{path}) {
+		t.Errorf("PipePaths = %q, want just %q", got, path)
+	}
+}

client/internal/daemonaddr/pipe_windows.go

@@ -0,0 +1,59 @@
+//go:build windows
+
+package daemonaddr
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net"
+
+	"github.com/Microsoft/go-winio"
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/sys/windows"
+
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
+)
+
+// dialPipePaths connects to the first path that answers with a pipe server this
+// client may trust, and returns the last error when none does.
+func dialPipePaths(ctx context.Context, paths []string) (net.Conn, error) {
+	var lastErr error
+	for _, path := range paths {
+		conn, err := dialPipe(ctx, path)
+		if err != nil {
+			log.Debugf("dial daemon pipe %s: %v", path, err)
+			lastErr = err
+			continue
+		}
+
+		// A pipe in the protected namespace could only have been created by an
+		// administrator or LocalSystem, so its name is the guarantee. Any other
+		// name has to be checked, because any local user can create one.
+		if !IsProtectedPipePath(path) {
+			if err := ipcauth.PipeServerTrusted(conn); err != nil {
+				if closeErr := conn.Close(); closeErr != nil {
+					log.Debugf("close untrusted pipe %s: %v", path, closeErr)
+				}
+				lastErr = fmt.Errorf("%s: %w", path, err)
+				continue
+			}
+		}
+
+		return conn, nil
+	}
+
+	if lastErr == nil {
+		lastErr = errors.New("no daemon pipe to connect to")
+	}
+	return nil, lastErr
+}
+
+// dialPipe connects to the daemon control pipe at SECURITY_IDENTIFICATION.
+// winio's plain DialPipe connects at SECURITY_ANONYMOUS, under which the daemon
+// cannot read the caller's token at all. Identification lets the daemon read the
+// caller's SID and groups without granting it the ability to act as the caller.
+func dialPipe(ctx context.Context, path string) (net.Conn, error) {
+	access := uint32(windows.GENERIC_READ | windows.GENERIC_WRITE)
+	return winio.DialPipeAccessImpLevel(ctx, path, access, winio.PipeImpLevelIdentification)
+}

client/internal/daemonaddr/resolve_pipe_other.go

@@ -0,0 +1,9 @@
+//go:build !windows
+
+package daemonaddr
+
+// ResolveDaemonAddr is a no-op off Windows, where there is no named-pipe
+// default to fall back from.
+func ResolveDaemonAddr(addr string) string {
+	return addr
+}

client/internal/daemonaddr/resolve_pipe_windows.go

@@ -0,0 +1,82 @@
+//go:build windows
+
+package daemonaddr
+
+import (
+	"net"
+	"strings"
+	"time"
+
+	"github.com/Microsoft/go-winio"
+	log "github.com/sirupsen/logrus"
+)
+
+// probeTimeout bounds each transport probe. Both are local, so a daemon that is
+// listening answers immediately and one that is not fails immediately.
+const probeTimeout = 300 * time.Millisecond
+
+// ResolveDaemonAddr keeps a client on the named pipe and never silently moves it
+// off. When the pipe does not answer it checks the legacy loopback TCP address, so
+// a client meeting a daemon that has not restarted since the upgrade can say what
+// is wrong, but it does not connect there.
+//
+// Using that address automatically would be a downgrade the user never asked for:
+// any local process can bind 127.0.0.1 while the daemon is not listening, and the
+// transport carries no caller identity, so a client that accepted whatever answered
+// would hand a setup key, a pre-shared key or an SSO prompt to a local impostor. An
+// operator who needs the legacy address during the upgrade window can still pass
+// --daemon-addr explicitly, which is a deliberate choice and still refuses the
+// privileged operations.
+//
+// Only the pipe address is resolved. A custom address is left alone, though passing
+// --daemon-addr npipe://netbird explicitly is indistinguishable from the default
+// here, so it is treated the same way.
+func ResolveDaemonAddr(addr string) string {
+	if addr != WindowsPipeAddr {
+		return addr
+	}
+
+	for _, path := range PipePaths("netbird") {
+		if pipeAvailable(path) {
+			return addr
+		}
+	}
+
+	if tcpAvailable(legacyWindowsAddr) {
+		log.Warnf("the daemon is not serving %s, but something is listening on the legacy %s. "+
+			"Restart the NetBird service so it serves the pipe. That address is not used automatically: "+
+			"any local user can bind it and it carries no caller identity, so pass --daemon-addr %s "+
+			"explicitly if you accept that",
+			WindowsPipeAddr, legacyWindowsAddr, legacyWindowsAddr)
+	}
+
+	return addr
+}
+
+func pipeAvailable(path string) bool {
+	timeout := probeTimeout
+	conn, err := winio.DialPipe(path, &timeout)
+	if err != nil {
+		return false
+	}
+	if err := conn.Close(); err != nil {
+		log.Debugf("close daemon pipe probe: %v", err)
+	}
+	return true
+}
+
+func tcpAvailable(addr string) bool {
+	host := addr
+	if _, after, ok := strings.Cut(addr, "://"); ok {
+		host = after
+	}
+
+	conn, err := net.DialTimeout("tcp", host, probeTimeout)
+	if err != nil {
+		return false
+	}
+	if err := conn.Close(); err != nil {
+		log.Debugf("close daemon TCP probe: %v", err)
+	}
+	return true
+}

client/internal/ipcauth/creds_stub.go

@@ -0,0 +1,31 @@
+//go:build !linux && !darwin && !freebsd && !windows
+
+package ipcauth
+
+import (
+	"errors"
+	"net"
+
+	"google.golang.org/grpc/credentials"
+)
+
+// errUnsupported is returned on platforms with no local peer-identity
+// primitive, so consumers fail closed instead of guessing an identity.
+var errUnsupported = errors.New("peer identity is not available on this platform")
+
+// NewTransportCredentials returns nil: without a peer-identity primitive the
+// daemon cannot authenticate local callers, and the caller must treat that as
+// "authorization cannot be enforced".
+func NewTransportCredentials() credentials.TransportCredentials {
+	return nil
+}
+
+// PeerIdentity always fails on this platform.
+func PeerIdentity(net.Conn) (Identity, error) {
+	return Identity{}, errUnsupported
+}
+
+// ConnIdentity always fails on this platform.
+func ConnIdentity(net.Conn) (Identity, error) {
+	return Identity{}, errUnsupported
+}

client/internal/ipcauth/creds_unix.go

@@ -0,0 +1,56 @@
+//go:build linux || darwin || freebsd
+
+package ipcauth
+
+import (
+	"context"
+	"net"
+
+	"google.golang.org/grpc/credentials"
+)
+
+// NewTransportCredentials returns gRPC transport credentials that expose the
+// caller's kernel-authenticated identity via IdentityFromContext. It returns
+// nil on platforms that have no peer-identity primitive, which the caller must
+// treat as "authorization cannot be enforced".
+//
+// The handshake exchanges no bytes on the wire, so a client dialing with
+// insecure credentials interoperates with a server using these. That keeps
+// older CLI and UI binaries working against an upgraded daemon.
+func NewTransportCredentials() credentials.TransportCredentials {
+	return unixCreds{}
+}
+
+// ConnIdentity extracts the caller's identity from an accepted local IPC
+// connection. It is shared by the gRPC transport credentials and by the JSON
+// gateway, which reads the identity of its own HTTP clients.
+func ConnIdentity(conn net.Conn) (Identity, error) {
+	return PeerIdentity(conn)
+}
+
+type unixCreds struct{}
+
+func (unixCreds) ClientHandshake(_ context.Context, _ string, conn net.Conn) (net.Conn, credentials.AuthInfo, error) {
+	return conn, AuthInfo{}, nil
+}
+
+// ServerHandshake extracts the peer identity and fails closed when it cannot
+// be read, so a connection whose caller is unknown never reaches a handler.
+func (unixCreds) ServerHandshake(conn net.Conn) (net.Conn, credentials.AuthInfo, error) {
+	id, err := ConnIdentity(conn)
+	if err != nil {
+		return nil, nil, err
+	}
+	return conn, AuthInfo{
+		CommonAuthInfo: credentials.CommonAuthInfo{SecurityLevel: credentials.NoSecurity},
+		Identity:       id,
+	}, nil
+}
+
+func (unixCreds) Info() credentials.ProtocolInfo {
+	return credentials.ProtocolInfo{SecurityProtocol: AuthInfo{}.AuthType()}
+}
+
+func (unixCreds) Clone() credentials.TransportCredentials { return unixCreds{} }
+
+func (unixCreds) OverrideServerName(string) error { return nil }

client/internal/ipcauth/creds_windows.go

@@ -0,0 +1,194 @@
+//go:build windows
+
+package ipcauth
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"runtime"
+
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/sys/windows"
+	"google.golang.org/grpc/credentials"
+)
+
+var (
+	modadvapi32                    = windows.NewLazySystemDLL("advapi32.dll")
+	procImpersonateNamedPipeClient = modadvapi32.NewProc("ImpersonateNamedPipeClient")
+)
+
+// DefaultPipeSDDL is the security descriptor for the daemon control pipe.
+//
+//	D:P          protected DACL, no inheritance
+//	(A;;GA;;;SY) allow GENERIC_ALL to LocalSystem (the daemon's service account)
+//	(A;;GA;;;WD) allow GENERIC_ALL to Everyone
+//
+// Any local caller may connect, as with a Unix socket at 0666; what a caller may
+// actually do is decided from its token, not from the DACL. Remote callers are not
+// a concern here: winio.ListenPipe creates the pipe with
+// FILE_PIPE_REJECT_REMOTE_CLIENTS, so NPFS rejects connections from other machines
+// before the descriptor is consulted.
+//
+// A deny ACE on the NETWORK SID would not add anything and would break callers:
+// that SID is present in any network-logon token, which includes OpenSSH and WinRM
+// sessions, so it denies administrators driving the CLI over SSH and denies the
+// daemon itself when started from such a session.
+func DefaultPipeSDDL() string {
+	return "D:P(A;;GA;;;SY)(A;;GA;;;WD)"
+}
+
+// NewTransportCredentials returns gRPC transport credentials that derive the
+// caller's identity from the named-pipe client token.
+//
+// The client must connect at SECURITY_IDENTIFICATION for the daemon to be able
+// to read its token, which is what DialNamedPipe does.
+func NewTransportCredentials() credentials.TransportCredentials {
+	return winpipeCreds{}
+}
+
+// ConnIdentity extracts the caller's identity from an accepted named-pipe
+// connection by impersonating the pipe client and reading its token. It is
+// shared by the gRPC transport credentials and by the JSON gateway, which
+// reads the identity of its own HTTP clients.
+func ConnIdentity(conn net.Conn) (Identity, error) {
+	// go-winio's pipe connection embeds *win32File, which exposes Fd().
+	fdConn, ok := conn.(interface{ Fd() uintptr })
+	if !ok {
+		return Identity{}, fmt.Errorf("connection %T does not expose a pipe handle", conn)
+	}
+	return pipeClientIdentity(windows.Handle(fdConn.Fd()))
+}
+
+type winpipeCreds struct{}
+
+func (winpipeCreds) ClientHandshake(_ context.Context, _ string, conn net.Conn) (net.Conn, credentials.AuthInfo, error) {
+	return conn, AuthInfo{}, nil
+}
+
+// ServerHandshake extracts the connecting client's identity and fails closed
+// when the handle or token cannot be read, so a connection whose caller is
+// unknown never reaches a handler.
+func (winpipeCreds) ServerHandshake(conn net.Conn) (net.Conn, credentials.AuthInfo, error) {
+	id, err := ConnIdentity(conn)
+	if err != nil {
+		return nil, nil, err
+	}
+	return conn, AuthInfo{
+		CommonAuthInfo: credentials.CommonAuthInfo{SecurityLevel: credentials.NoSecurity},
+		Identity:       id,
+	}, nil
+}
+
+func (winpipeCreds) Info() credentials.ProtocolInfo {
+	return credentials.ProtocolInfo{SecurityProtocol: AuthInfo{}.AuthType()}
+}
+
+func (winpipeCreds) Clone() credentials.TransportCredentials { return winpipeCreds{} }
+
+func (winpipeCreds) OverrideServerName(string) error { return nil }
+
+// pipeClientIdentity reads the connecting client's user SID, usable group
+// SIDs, and elevation state by impersonating the pipe client on this thread
+// and reading the resulting impersonation token.
+func pipeClientIdentity(handle windows.Handle) (id Identity, err error) {
+	// Impersonation is per-thread, so the goroutine must stay on this thread
+	// until RevertToSelf, otherwise an unrelated goroutine could inherit the
+	// impersonated context.
+	runtime.LockOSThread()
+
+	// The thread only goes back to the runtime's pool once it is provably no
+	// longer impersonating the client. If the revert fails, leaving it locked
+	// makes Go terminate it when this goroutine exits, which costs one thread
+	// and keeps a thread running as the client from ever being reused.
+	clean := false
+	defer func() {
+		if clean {
+			runtime.UnlockOSThread()
+		}
+	}()
+
+	if err = impersonateNamedPipeClient(handle); err != nil {
+		clean = true
+		return Identity{}, fmt.Errorf("impersonate named pipe client: %w", err)
+	}
+	defer func() {
+		// Surface the revert failure only when nothing else failed: leaving
+		// the thread impersonated is worse than the original error.
+		revErr := windows.RevertToSelf()
+		if revErr != nil {
+			if err == nil {
+				err = fmt.Errorf("revert impersonation: %w", revErr)
+			}
+			return
+		}
+		clean = true
+	}()
+
+	// openAsSelf=true opens the token with the daemon's own process context
+	// rather than the impersonated client's, so the open cannot fail because
+	// the client lacks access to its own token.
+	var token windows.Token
+	if err = windows.OpenThreadToken(windows.CurrentThread(), windows.TOKEN_QUERY, true, &token); err != nil {
+		return Identity{}, fmt.Errorf("open thread token: %w", err)
+	}
+	defer func() {
+		if cerr := token.Close(); cerr != nil {
+			log.Debugf("close client token: %v", cerr)
+		}
+	}()
+
+	return identityFromToken(token)
+}
+
+// identityFromToken reads the user SID, usable group SIDs and elevation state
+// out of a Windows token.
+func identityFromToken(token windows.Token) (Identity, error) {
+	user, err := token.GetTokenUser()
+	if err != nil {
+		return Identity{}, fmt.Errorf("read token user: %w", err)
+	}
+
+	groups, err := tokenGroupSIDs(token)
+	if err != nil {
+		return Identity{}, err
+	}
+
+	return Identity{
+		SID:      user.User.Sid.String(),
+		Groups:   groups,
+		Elevated: token.IsElevated(),
+	}, nil
+}
+
+// tokenGroupSIDs returns the SIDs of the groups the token can actually
+// exercise. Groups that are disabled or marked deny-only are skipped: a
+// UAC-filtered administrator carries BUILTIN\Administrators as deny-only, and
+// treating that as membership would hand every admin account privilege it
+// cannot currently use.
+func tokenGroupSIDs(token windows.Token) ([]string, error) {
+	tg, err := token.GetTokenGroups()
+	if err != nil {
+		return nil, fmt.Errorf("read token groups: %w", err)
+	}
+
+	var sids []string
+	for _, g := range tg.AllGroups() {
+		if g.Attributes&windows.SE_GROUP_ENABLED == 0 {
+			continue
+		}
+		if g.Attributes&windows.SE_GROUP_USE_FOR_DENY_ONLY != 0 {
+			continue
+		}
+		sids = append(sids, g.Sid.String())
+	}
+	return sids, nil
+}
+
+func impersonateNamedPipeClient(h windows.Handle) error {
+	r, _, e := procImpersonateNamedPipeClient.Call(uintptr(h))
+	if r == 0 {
+		return e
+	}
+	return nil
+}

client/internal/ipcauth/forward.go

@@ -0,0 +1,272 @@
+package ipcauth
+
+import (
+	"context"
+	"crypto/rand"
+	"crypto/subtle"
+	"encoding/hex"
+	"fmt"
+	"slices"
+	"strconv"
+	"strings"
+
+	"google.golang.org/grpc/metadata"
+)
+
+// Metadata keys the local JSON gateway uses to forward the identity of its own
+// HTTP client to the daemon. The gateway runs inside the daemon process and
+// re-dials the daemon over the control socket, so without forwarding every
+// JSON request would appear to come from the daemon itself.
+const (
+	// mdFwd marks a request as forwarded by the JSON gateway. It is always
+	// set, even when the gateway could not read its client's identity, so the
+	// daemon can tell "no identity available" apart from "not forwarded".
+	mdFwd         = "x-netbird-fwd"
+	mdFwdUID      = "x-netbird-fwd-uid"      // Unix user ID
+	mdFwdGID      = "x-netbird-fwd-gid"      // Unix primary group ID
+	mdFwdSID      = "x-netbird-fwd-sid"      // Windows user SID
+	mdFwdGroup    = "x-netbird-fwd-group"    // Windows group SID, repeated
+	mdFwdElevated = "x-netbird-fwd-elevated" // Windows, "1" when elevated
+
+	// mdFwdProof proves the forwarded identity was stamped by this process. The
+	// gateway runs inside the daemon, so a secret held in memory is available to
+	// the only legitimate producer and to nothing else.
+	mdFwdProof = "x-netbird-fwd-proof"
+)
+
+// forwardKeys is every metadata key the gateway sets. An HTTP client must never
+// be able to supply one itself: see IsReservedForwardKey.
+var forwardKeys = []string{mdFwd, mdFwdUID, mdFwdGID, mdFwdSID, mdFwdGroup, mdFwdElevated, mdFwdProof}
+
+// forwardProof authenticates the gateway's forwarding metadata. It is generated
+// once per daemon process and never leaves it: it is not written to disk, not
+// logged, and not sent anywhere except over the daemon's own control socket to
+// itself.
+//
+// Without it, trusting a forwarded identity rests on every layer in front of it
+// stripping incoming forwarding keys, and on each key's value shape being
+// distinguishable from an injected one. A single injected group SID or an
+// injected "elevated" flag has the same shape as a legitimate one, so no
+// cardinality rule can catch it. Requiring the proof means metadata that did not
+// come from this process is refused whatever it contains.
+var forwardProof = mustForwardProof()
+
+func mustForwardProof() string {
+	var buf [32]byte
+	if _, err := rand.Read(buf[:]); err != nil {
+		// Continuing would leave the forwarded path authenticated by a
+		// predictable value, which is worse than not starting.
+		panic(fmt.Sprintf("generate identity forwarding proof: %v", err))
+	}
+	return hex.EncodeToString(buf[:])
+}
+
+// IsReservedForwardKey reports whether a gRPC metadata key belongs to the
+// gateway's identity forwarding, and therefore must be dropped when it arrives
+// from outside.
+//
+// grpc-gateway maps "Grpc-Metadata-<key>" request headers into gRPC metadata and
+// joins them ahead of the values its own annotators add. Without dropping these,
+// an HTTP client could hand the daemon "x-netbird-fwd-uid: 0" and be believed,
+// because the daemon trusts forwarded metadata when the transport peer is the
+// (privileged) gateway.
+func IsReservedForwardKey(key string) bool {
+	key = strings.ToLower(key)
+	return slices.Contains(forwardKeys, key)
+}
+
+// ForwardIdentityMetadata encodes an HTTP client's identity for the JSON
+// gateway to forward to the daemon. When known is false only the marker is
+// set, which makes the daemon treat the caller as unidentified rather than as
+// the daemon itself.
+func ForwardIdentityMetadata(id Identity, known bool) metadata.MD {
+	md := metadata.MD{}
+	md.Set(mdFwd, "1")
+	md.Set(mdFwdProof, forwardProof)
+	if !known {
+		return md
+	}
+
+	if id.IsWindows() {
+		md.Set(mdFwdSID, id.SID)
+		if len(id.Groups) > 0 {
+			md.Set(mdFwdGroup, id.Groups...)
+		}
+		if id.Elevated {
+			md.Set(mdFwdElevated, "1")
+		}
+		return md
+	}
+
+	md.Set(mdFwdUID, strconv.FormatUint(uint64(id.UID), 10))
+	md.Set(mdFwdGID, strconv.FormatUint(uint64(id.GID), 10))
+	return md
+}
+
+// CallerIdentity returns the identity to authorize a request against. For a
+// direct connection that is the transport peer's kernel identity. For a
+// request relayed by the local JSON gateway it is the identity the gateway
+// forwarded, since the transport peer is then the daemon itself.
+//
+// A forwarded identity is only honoured when the transport peer is the daemon's
+// own identity and the metadata carries this process's forwarding proof, so
+// forged forwarding metadata gains a caller nothing. A forwarded request that
+// carries no identity is reported as unidentified, never as the daemon.
+//
+// The second return value is false when no identity could be established, and
+// callers MUST fail closed in that case.
+func CallerIdentity(ctx context.Context) (Identity, bool) {
+	id, ok := IdentityFromContext(ctx)
+	if !ok {
+		return Identity{}, false
+	}
+
+	// A forwarding key that arrives more than once did not come from the gateway
+	// alone, so nothing about the request can be trusted to describe its caller.
+	// Refusing outright matters because the alternative reading, "not forwarded",
+	// would authorize the request as the transport peer, which on the gateway's
+	// connection is the daemon itself.
+	if duplicatedForwardKey(ctx) {
+		return Identity{}, false
+	}
+
+	forwarded := isForwarded(ctx)
+
+	// Our own process on the other end of the socket is the JSON gateway, the only
+	// thing that dials the daemon from inside it. Such a call must carry a
+	// forwarded identity; without one there is no caller to authorize, and
+	// treating it as the daemon would authorize whatever reached the JSON socket.
+	// Only Linux reports the peer PID, so this is a belt on top of the gateway's
+	// interceptor rather than the sole guarantee.
+	if id.PID != 0 && int(id.PID) == selfPID && !forwarded {
+		return Identity{}, false
+	}
+
+	// Only the gateway's own connection may speak for someone else. Being
+	// privileged is not enough and not the point: the gateway runs inside the
+	// daemon, so it dials as the daemon's identity whatever user that is, which
+	// also covers a rootless container.
+	if !forwarded || !IsDaemonSelf(id) {
+		return id, true
+	}
+
+	// Speaking for someone else additionally requires the proof only this process
+	// holds. Refusing is the only safe reading: the transport peer here is the
+	// daemon itself, so falling back to it would authorize the request as the
+	// daemon. This is also what makes the forwarded values trustworthy once
+	// accepted, so they need no shape checks of their own.
+	if !authenticForward(ctx) {
+		return Identity{}, false
+	}
+
+	return forwardedIdentity(ctx)
+}
+
+// duplicatedForwardKey reports whether any forwarding key carries more than one
+// value. The gateway's interceptor sets each key exactly once and replaces what
+// was already there, so a repeat means a second source supplied it.
+func duplicatedForwardKey(ctx context.Context) bool {
+	md, ok := metadata.FromIncomingContext(ctx)
+	if !ok {
+		return false
+	}
+	for _, key := range forwardKeys {
+		// Group SIDs are legitimately repeated; the rest identify the caller.
+		if key == mdFwdGroup {
+			continue
+		}
+		if len(md.Get(key)) > 1 {
+			return true
+		}
+	}
+	return false
+}
+
+// authenticForward reports whether the request carries this process's forwarding
+// proof, which only the in-process JSON gateway can supply.
+func authenticForward(ctx context.Context) bool {
+	md, ok := metadata.FromIncomingContext(ctx)
+	if !ok {
+		return false
+	}
+	got := mdSingle(md, mdFwdProof)
+	return subtle.ConstantTimeCompare([]byte(got), []byte(forwardProof)) == 1
+}
+
+// isForwarded reports whether the request carries the JSON gateway marker.
+func isForwarded(ctx context.Context) bool {
+	md, ok := metadata.FromIncomingContext(ctx)
+	if !ok {
+		return false
+	}
+	return mdSingle(md, mdFwd) != ""
+}
+
+// forwardedIdentity decodes the identity the JSON gateway attached.
+func forwardedIdentity(ctx context.Context) (Identity, bool) {
+	md, ok := metadata.FromIncomingContext(ctx)
+	if !ok {
+		return Identity{}, false
+	}
+
+	if sid := mdSingle(md, mdFwdSID); sid != "" {
+		return Identity{
+			SID: sid,
+			// Repeated by design, one value per group, and only reachable once
+			// the forwarding proof has been verified.
+			Groups:   md.Get(mdFwdGroup),
+			Elevated: mdSingle(md, mdFwdElevated) == "1",
+		}, true
+	}
+
+	uid, err := strconv.ParseUint(mdSingle(md, mdFwdUID), 10, 32)
+	if err != nil {
+		return Identity{}, false
+	}
+
+	id := Identity{UID: uint32(uid)}
+	if gid, err := strconv.ParseUint(mdSingle(md, mdFwdGID), 10, 32); err == nil {
+		id.GID = uint32(gid)
+	}
+	return id, true
+}
+
+// mdSingle returns the value of a forwarded key only when exactly one was
+// supplied. The gateway's interceptor sets each key exactly once, so more than one
+// value means something else also supplied it, and the whole identity is treated as
+// unknown rather than picking a winner. Defence in depth behind the gateway's
+// header filter.
+func mdSingle(md metadata.MD, key string) string {
+	if v := md.Get(key); len(v) == 1 {
+		return v[0]
+	}
+	return ""
+}
+
+// WithForwardedIdentity stamps id onto a context's outgoing metadata for the JSON
+// gateway's call to the daemon, replacing any forwarding keys already present so
+// values supplied from outside cannot survive alongside it.
+//
+// This is deliberately not done with runtime.WithMetadata: grpc-gateway skips its
+// annotators entirely when no request header maps to metadata ("if len(pairs) == 0
+// { return ctx, nil, nil }", runtime/context.go), which an HTTP/1.0 request with no
+// Host header over a unix socket achieves. The daemon would then see an unmarked
+// call whose transport peer is the daemon's own identity, and authorize it as the
+// daemon. A client interceptor runs for every RPC regardless of headers.
+func WithForwardedIdentity(ctx context.Context, id Identity, known bool) context.Context {
+	md, ok := metadata.FromOutgoingContext(ctx)
+	if !ok {
+		md = metadata.MD{}
+	} else {
+		md = md.Copy()
+	}
+
+	for _, key := range forwardKeys {
+		delete(md, key)
+	}
+	for key, values := range ForwardIdentityMetadata(id, known) {
+		md[key] = values
+	}
+
+	return metadata.NewOutgoingContext(ctx, md)
+}

client/internal/ipcauth/forward_test.go

@@ -0,0 +1,214 @@
+package ipcauth
+
+import (
+	"context"
+	"testing"
+
+	"google.golang.org/grpc/credentials"
+	"google.golang.org/grpc/metadata"
+	"google.golang.org/grpc/peer"
+)
+
+// transportCtx builds a request context as the daemon's transport credentials
+// would: the identity of whoever opened the socket, plus whatever metadata the
+// request carried.
+func transportCtx(id Identity, md metadata.MD) context.Context {
+	ctx := peer.NewContext(context.Background(), &peer.Peer{
+		AuthInfo: AuthInfo{
+			CommonAuthInfo: credentials.CommonAuthInfo{SecurityLevel: credentials.NoSecurity},
+			Identity:       id,
+		},
+	})
+	if md != nil {
+		ctx = metadata.NewIncomingContext(ctx, md)
+	}
+	return ctx
+}
+
+var (
+	root       = Identity{UID: 0}
+	unprivUser = Identity{UID: 1000, GID: 1000}
+)
+
+// asDaemon pins which identity counts as this process for the duration of a test.
+// Without it the test binary's own uid decides, which silently changes what
+// "the gateway" means.
+func asDaemon(t *testing.T, id Identity) {
+	t.Helper()
+	prevID, prevKnown, prevDelegate := selfIdentity, selfKnown, selfMayDelegate
+	t.Cleanup(func() { selfIdentity, selfKnown, selfMayDelegate = prevID, prevKnown, prevDelegate })
+	selfIdentity, selfKnown = id, true
+	selfMayDelegate = !id.IsPrivileged()
+}
+
+func TestCallerIdentity_DirectConnections(t *testing.T) {
+	t.Run("no transport credentials is not an identity", func(t *testing.T) {
+		if _, ok := CallerIdentity(context.Background()); ok {
+			t.Fatal("a caller with no credentials must not be identified")
+		}
+	})
+
+	t.Run("a direct caller is its transport identity", func(t *testing.T) {
+		id, ok := CallerIdentity(transportCtx(unprivUser, nil))
+		if !ok || id.UID != 1000 {
+			t.Fatalf("got %v ok=%t, want uid 1000", id, ok)
+		}
+	})
+
+	// The whole point of honouring forwarded metadata only from a privileged
+	// transport peer: an unprivileged caller can set any metadata it likes on its
+	// own connection to the daemon socket.
+	t.Run("an unprivileged caller cannot forge an identity", func(t *testing.T) {
+		asDaemon(t, root)
+		forged := metadata.Pairs(mdFwd, "1", mdFwdUID, "0", mdFwdGID, "0")
+		id, ok := CallerIdentity(transportCtx(unprivUser, forged))
+		if !ok {
+			t.Fatal("caller should still be identified, as itself")
+		}
+		if id.IsPrivileged() || id.UID != 1000 {
+			t.Fatalf("forged metadata was believed: got %v", id)
+		}
+	})
+}
+
+func TestCallerIdentity_GatewayForwarding(t *testing.T) {
+	t.Run("the gateway's client identity is used, not the gateway's own", func(t *testing.T) {
+		asDaemon(t, root)
+		md := ForwardIdentityMetadata(unprivUser, true)
+		id, ok := CallerIdentity(transportCtx(root, md))
+		if !ok {
+			t.Fatal("forwarded identity should be usable")
+		}
+		if id.IsPrivileged() || id.UID != 1000 {
+			t.Fatalf("got %v, want the forwarded uid 1000 and not privileged", id)
+		}
+	})
+
+	t.Run("a privileged gateway client stays privileged", func(t *testing.T) {
+		asDaemon(t, root)
+		md := ForwardIdentityMetadata(root, true)
+		id, ok := CallerIdentity(transportCtx(root, md))
+		if !ok || !id.IsPrivileged() {
+			t.Fatalf("got %v ok=%t, want a privileged identity", id, ok)
+		}
+	})
+
+	// A JSON socket the gateway cannot read peer credentials from (a TCP socket,
+	// say) must not make every request look like the daemon itself.
+	t.Run("an unreadable client identity is unknown, not the daemon", func(t *testing.T) {
+		asDaemon(t, root)
+		md := ForwardIdentityMetadata(Identity{}, false)
+		if _, ok := CallerIdentity(transportCtx(root, md)); ok {
+			t.Fatal("a forwarded request with no identity must not be identified")
+		}
+	})
+
+	// grpc-gateway turns Grpc-Metadata-<key> headers into gRPC metadata and joins
+	// them ahead of its annotators' values. If an HTTP client's header survived
+	// that, this is the shape the daemon would see: the attacker's uid 0 first,
+	// the real uid second. The gateway filters those headers out, and reading a
+	// duplicated key as unknown makes the daemon safe even if it did not.
+	t.Run("a duplicated key from an injected header is not believed", func(t *testing.T) {
+		asDaemon(t, root)
+		md := metadata.MD{}
+		md.Append(mdFwd, "1")
+		md.Append(mdFwdUID, "0")    // injected by the HTTP client
+		md.Append(mdFwdUID, "1000") // appended by the gateway's annotator
+		if id, ok := CallerIdentity(transportCtx(root, md)); ok {
+			t.Fatalf("injected uid was accepted: got %v", id)
+		}
+	})
+
+	t.Run("a duplicated marker is not believed either", func(t *testing.T) {
+		asDaemon(t, root)
+		md := metadata.MD{}
+		md.Append(mdFwd, "1")
+		md.Append(mdFwd, "1")
+		md.Append(mdFwdUID, "1000")
+		// A repeated marker must not be read as "not forwarded": that would
+		// authorize the request as the transport peer, which on the gateway's
+		// connection is the daemon itself.
+		if id, ok := CallerIdentity(transportCtx(root, md)); ok {
+			t.Fatalf("a duplicated marker was believed: got %v", id)
+		}
+	})
+
+	// The layers in front of this (the gateway's header matcher, and its
+	// interceptor replacing every forwarding key) are what keep outside metadata
+	// from arriving at all. The proof is what the daemon can check for itself, and
+	// it is the only defence that works for a value whose legitimate shape is
+	// indistinguishable from an injected one: a lone group SID, or "elevated".
+	t.Run("forwarding metadata without this process's proof is refused", func(t *testing.T) {
+		asDaemon(t, root)
+		for name, md := range map[string]metadata.MD{
+			"no proof":    metadata.Pairs(mdFwd, "1", mdFwdUID, "0"),
+			"wrong proof": metadata.Pairs(mdFwd, "1", mdFwdUID, "0", mdFwdProof, "deadbeef"),
+			"windows identity without a proof": metadata.Pairs(mdFwd, "1",
+				mdFwdSID, "S-1-5-21-1-2-3-1001", mdFwdGroup, sidAdministrators, mdFwdElevated, "1"),
+		} {
+			t.Run(name, func(t *testing.T) {
+				if id, ok := CallerIdentity(transportCtx(root, md)); ok {
+					t.Fatalf("unstamped forwarding metadata was believed: got %v", id)
+				}
+			})
+		}
+	})
+
+	// A caller that reaches the gateway cannot see the proof, so it cannot append
+	// a group of its own to a genuine forwarded identity: doing so would have to
+	// go through the interceptor, which replaces the whole set.
+	t.Run("a group appended to a stamped identity does not survive the interceptor", func(t *testing.T) {
+		asDaemon(t, root)
+		injected := metadata.MD{}
+		injected.Append(mdFwdGroup, sidAdministrators)
+
+		ctx := WithForwardedIdentity(metadata.NewOutgoingContext(context.Background(), injected),
+			Identity{SID: "S-1-5-21-1-2-3-1001"}, true)
+		out, ok := metadata.FromOutgoingContext(ctx)
+		if !ok {
+			t.Fatal("no outgoing metadata")
+		}
+		if groups := out.Get(mdFwdGroup); len(groups) != 0 {
+			t.Fatalf("injected group survived: %v", groups)
+		}
+	})
+}
+
+func TestIsReservedForwardKey(t *testing.T) {
+	for _, key := range forwardKeys {
+		if !IsReservedForwardKey(key) {
+			t.Errorf("%q must be reserved", key)
+		}
+	}
+
+	// grpc-gateway canonicalises header names, so the check has to be
+	// case-insensitive.
+	if !IsReservedForwardKey("X-Netbird-Fwd-Uid") {
+		t.Error("the check must be case-insensitive")
+	}
+
+	for _, key := range []string{"authorization", "x-netbird", "x-netbird-fwd-uid-extra", ""} {
+		if IsReservedForwardKey(key) {
+			t.Errorf("%q must not be reserved", key)
+		}
+	}
+}
+
+func TestForwardIdentityMetadata_AlwaysMarksForwarded(t *testing.T) {
+	for _, tc := range []struct {
+		name  string
+		id    Identity
+		known bool
+	}{
+		{"known unix identity", unprivUser, true},
+		{"unknown identity", Identity{}, false},
+		{"windows identity", Identity{SID: "S-1-5-21-1-2-3-1001", Elevated: true}, true},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			md := ForwardIdentityMetadata(tc.id, tc.known)
+			if got := md.Get(mdFwd); len(got) != 1 || got[0] != "1" {
+				t.Fatalf("marker = %v, want exactly one \"1\"", got)
+			}
+		})
+	}
+}

client/internal/ipcauth/identity.go

@@ -0,0 +1,127 @@
+// Package ipcauth provides the kernel-authenticated identity of a local IPC
+// (gRPC) caller and the transport credentials that surface it into the gRPC
+// context, so the daemon can authorize individual RPCs by caller identity.
+//
+// On Unix the identity is read from the kernel via SO_PEERCRED (Linux) or
+// LOCAL_PEERCRED (Darwin/FreeBSD). On Windows it is derived from the
+// named-pipe client token. Platforms without a peer-identity primitive get no
+// credentials, and every consumer must fail closed when no identity is
+// available.
+package ipcauth
+
+import (
+	"context"
+	"fmt"
+	"slices"
+
+	"google.golang.org/grpc/credentials"
+	"google.golang.org/grpc/peer"
+)
+
+// Well-known Windows SIDs that identify a fully privileged principal.
+const (
+	sidLocalSystem    = "S-1-5-18"     // NT AUTHORITY\SYSTEM
+	sidLocalService   = "S-1-5-19"     // NT AUTHORITY\LOCAL SERVICE
+	sidNetworkService = "S-1-5-20"     // NT AUTHORITY\NETWORK SERVICE
+	sidAdministrators = "S-1-5-32-544" // BUILTIN\Administrators
+)
+
+// Identity is the kernel-authenticated identity of a local IPC caller. The
+// zero value is not a valid identity: consumers must only use one obtained
+// with a true ok/nil error return.
+type Identity struct {
+	// UID and GID are the caller's Unix user ID and primary group ID. Both are
+	// zero on Windows, where SID is authoritative instead.
+	UID uint32
+	GID uint32
+
+	// SID is the caller's Windows security identifier, empty on Unix.
+	SID string
+
+	// Groups holds the caller's Windows group SIDs, captured from the client
+	// token at handshake time. Only groups that are enabled and not
+	// deny-only are captured, so a group listed here is one the caller can
+	// actually exercise. Empty on Unix.
+	Groups []string
+
+	// Elevated reports whether the Windows client token is elevated (running
+	// as administrator, or an administrator with UAC turned off). Always false
+	// on Unix, where privilege is uid 0.
+	Elevated bool
+
+	// PID is the caller's process ID where the platform reports it (Linux's
+	// SO_PEERCRED), and 0 where it does not. It identifies the daemon's own
+	// process dialling itself, which is what the JSON gateway does, and is never
+	// used to grant anything.
+	PID int32
+}
+
+// IsWindows reports whether this identity is a Windows principal (SID-based)
+// rather than a Unix uid/gid principal.
+func (i Identity) IsWindows() bool {
+	return i.SID != ""
+}
+
+// IsPrivileged reports whether the caller is the platform's administrative
+// principal, which is what the daemon requires for changes that cross the
+// user-to-root boundary.
+//
+// On Windows the decision comes from the caller's token rather than from
+// account names or group RIDs: an elevated token, one of the service accounts
+// the daemon itself may run as, or a token with BUILTIN\Administrators
+// enabled. A UAC-filtered administrator has that group marked deny-only, and
+// deny-only groups are dropped when the identity is captured, so such a
+// caller is correctly reported as unprivileged. Domain group memberships
+// (Domain Admins and friends) are deliberately not consulted: they say
+// nothing about what this token may do on this machine.
+func (i Identity) IsPrivileged() bool {
+	if !i.IsWindows() {
+		return i.UID == 0
+	}
+
+	if i.Elevated {
+		return true
+	}
+
+	switch i.SID {
+	case sidLocalSystem, sidLocalService, sidNetworkService:
+		return true
+	}
+
+	return slices.Contains(i.Groups, sidAdministrators)
+}
+
+// String renders the identity for audit logs and denial messages.
+func (i Identity) String() string {
+	if i.IsWindows() {
+		return fmt.Sprintf("sid=%s elevated=%t", i.SID, i.Elevated)
+	}
+	return fmt.Sprintf("uid=%d gid=%d", i.UID, i.GID)
+}
+
+// AuthInfo carries the peer Identity as a gRPC credentials.AuthInfo so
+// handlers can retrieve it from the request context via IdentityFromContext.
+type AuthInfo struct {
+	credentials.CommonAuthInfo
+	Identity Identity
+}
+
+// AuthType identifies the authentication scheme.
+func (AuthInfo) AuthType() string { return "netbird-ipc-peercred" }
+
+// IdentityFromContext extracts the caller's kernel-authenticated identity from
+// the gRPC peer context. The second return value is false when no IPC
+// transport credentials were negotiated, which happens on a TCP daemon socket
+// and on platforms without a peer-identity primitive. Callers MUST fail closed
+// in that case.
+func IdentityFromContext(ctx context.Context) (Identity, bool) {
+	p, ok := peer.FromContext(ctx)
+	if !ok {
+		return Identity{}, false
+	}
+	info, ok := p.AuthInfo.(AuthInfo)
+	if !ok {
+		return Identity{}, false
+	}
+	return info.Identity, true
+}

client/internal/ipcauth/peercred_bsd.go

@@ -0,0 +1,43 @@
+//go:build darwin || freebsd
+
+package ipcauth
+
+import (
+	"fmt"
+	"net"
+
+	"golang.org/x/sys/unix"
+)
+
+// PeerIdentity reads the kernel-authenticated identity of the process on the
+// other end of a Unix socket via LOCAL_PEERCRED. The xucred is recorded by the
+// kernel at connect() time and carries the peer's uid and its group list, of
+// which the first entry is the primary group.
+func PeerIdentity(conn net.Conn) (Identity, error) {
+	uc, ok := conn.(*net.UnixConn)
+	if !ok {
+		return Identity{}, fmt.Errorf("connection is not a unix socket: %T", conn)
+	}
+
+	raw, err := uc.SyscallConn()
+	if err != nil {
+		return Identity{}, fmt.Errorf("raw conn: %w", err)
+	}
+
+	var cred *unix.Xucred
+	var credErr error
+	if err := raw.Control(func(fd uintptr) {
+		cred, credErr = unix.GetsockoptXucred(int(fd), unix.SOL_LOCAL, unix.LOCAL_PEERCRED)
+	}); err != nil {
+		return Identity{}, fmt.Errorf("control raw conn: %w", err)
+	}
+	if credErr != nil {
+		return Identity{}, fmt.Errorf("read LOCAL_PEERCRED: %w", credErr)
+	}
+
+	id := Identity{UID: cred.Uid}
+	if cred.Ngroups > 0 {
+		id.GID = cred.Groups[0]
+	}
+	return id, nil
+}

client/internal/ipcauth/peercred_linux.go

@@ -0,0 +1,39 @@
+//go:build linux
+
+package ipcauth
+
+import (
+	"fmt"
+	"net"
+
+	"golang.org/x/sys/unix"
+)
+
+// PeerIdentity reads the kernel-authenticated identity of the process on the
+// other end of a Unix socket via SO_PEERCRED. The credentials are recorded by
+// the kernel at connect() time and cannot be changed for the life of the
+// connection, so they are not spoofable by the caller.
+func PeerIdentity(conn net.Conn) (Identity, error) {
+	uc, ok := conn.(*net.UnixConn)
+	if !ok {
+		return Identity{}, fmt.Errorf("connection is not a unix socket: %T", conn)
+	}
+
+	raw, err := uc.SyscallConn()
+	if err != nil {
+		return Identity{}, fmt.Errorf("raw conn: %w", err)
+	}
+
+	var cred *unix.Ucred
+	var credErr error
+	if err := raw.Control(func(fd uintptr) {
+		cred, credErr = unix.GetsockoptUcred(int(fd), unix.SOL_SOCKET, unix.SO_PEERCRED)
+	}); err != nil {
+		return Identity{}, fmt.Errorf("control raw conn: %w", err)
+	}
+	if credErr != nil {
+		return Identity{}, fmt.Errorf("read SO_PEERCRED: %w", credErr)
+	}
+
+	return Identity{UID: cred.Uid, GID: cred.Gid, PID: cred.Pid}, nil
+}

client/internal/ipcauth/pipeserver_windows.go

@@ -0,0 +1,87 @@
+//go:build windows
+
+package ipcauth
+
+import (
+	"fmt"
+	"net"
+
+	log "github.com/sirupsen/logrus"
+	"golang.org/x/sys/windows"
+)
+
+// PipeServerTrusted reports an error unless the pipe behind conn was created by a
+// principal this client may hand secrets to. Clients call it for a pipe whose name
+// carries no guarantee of its own, which is any name outside the
+// ProtectedPrefix\Administrators namespace: that namespace already restricts
+// creation to administrators and LocalSystem, while a plain name can be created by
+// any local user before the daemon gets there.
+//
+// The decision is made from the pipe object's owner, not from the serving process,
+// because a client cannot open a process running as another user at all, and the
+// legitimate case is precisely an unprivileged client talking to a privileged
+// daemon. Trusted owners are the service accounts, BUILTIN\Administrators, and
+// this client's own user, the last of which is the daemon a user runs themselves
+// as in netstack mode. A pipe owned by anyone else gets no setup key, pre-shared
+// key or SSO prompt out of this client.
+func PipeServerTrusted(conn net.Conn) error {
+	// go-winio's pipe connection embeds *win32File, which exposes Fd().
+	fdConn, ok := conn.(interface{ Fd() uintptr })
+	if !ok {
+		return fmt.Errorf("connection %T does not expose a pipe handle", conn)
+	}
+
+	owner, err := pipeOwnerSID(windows.Handle(fdConn.Fd()))
+	if err != nil {
+		return err
+	}
+
+	if !trustedPipeOwner(owner) {
+		return fmt.Errorf("pipe owned by %s, which is neither an administrator nor this user", owner)
+	}
+	return nil
+}
+
+// PipeOwnedBySelf reports whether the pipe behind conn was created by this very
+// user, which is how a client recognises a daemon running as itself. Ownership it
+// cannot read is reported as false.
+func PipeOwnedBySelf(conn net.Conn) bool {
+	fdConn, ok := conn.(interface{ Fd() uintptr })
+	if !ok {
+		return false
+	}
+
+	owner, err := pipeOwnerSID(windows.Handle(fdConn.Fd()))
+	if err != nil {
+		log.Debugf("read daemon pipe owner: %v", err)
+		return false
+	}
+	return selfKnown && selfIdentity.SID != "" && owner == selfIdentity.SID
+}
+
+// pipeOwnerSID reads the owner of the pipe object a client is connected to. The
+// handle was opened with GENERIC_READ, which includes READ_CONTROL, so no extra
+// access is needed.
+func pipeOwnerSID(handle windows.Handle) (string, error) {
+	sd, err := windows.GetSecurityInfo(handle, windows.SE_KERNEL_OBJECT, windows.OWNER_SECURITY_INFORMATION)
+	if err != nil {
+		return "", fmt.Errorf("read pipe security info: %w", err)
+	}
+
+	owner, _, err := sd.Owner()
+	if err != nil {
+		return "", fmt.Errorf("read pipe owner: %w", err)
+	}
+	return owner.String(), nil
+}
+
+// trustedPipeOwner reports whether a pipe's owner is a principal a client may
+// speak to. An elevated process's objects are owned by BUILTIN\Administrators by
+// default, an unelevated one's by the user, which is why both forms appear here.
+func trustedPipeOwner(owner string) bool {
+	switch owner {
+	case sidLocalSystem, sidLocalService, sidNetworkService, sidAdministrators:
+		return true
+	}
+	return selfKnown && selfIdentity.SID != "" && owner == selfIdentity.SID
+}

client/internal/ipcauth/privileged.go

@@ -0,0 +1,125 @@
+package ipcauth
+
+import (
+	"os"
+	"runtime"
+)
+
+// Fields of the ErrorInfo detail the daemon attaches to a PermissionDenied it
+// raises for an operation that requires root/administrator. Clients match on
+// Reason and Domain rather than on the message text, and render the summary and
+// command themselves so the user gets guidance instead of a gRPC error dump.
+const (
+	// ErrorReasonPrivilegeRequired identifies the detail.
+	ErrorReasonPrivilegeRequired = "PRIVILEGE_REQUIRED"
+	// ErrorDomain scopes the reason to the NetBird daemon.
+	ErrorDomain = "daemon.netbird.io"
+	// ErrorMetaSummary is the one-sentence explanation of what was refused.
+	ErrorMetaSummary = "summary"
+	// ErrorMetaCommand is the command that performs the same operation with the
+	// privileges it needs, ready to copy and run.
+	ErrorMetaCommand = "command"
+)
+
+// The identity of the process evaluating callers, captured once because it cannot
+// change. selfKnown is false when it could not be read, in which case nothing is
+// ever treated as this process. selfMayDelegate additionally requires this
+// process to be unprivileged: see IsPrivilegedCaller.
+var (
+	selfIdentity    Identity
+	selfKnown       bool
+	selfMayDelegate bool
+	// selfPID is this process's PID, used to recognise the daemon dialling itself.
+	selfPID = os.Getpid()
+)
+
+func init() {
+	id, err := CurrentProcessIdentity()
+	if err != nil {
+		return
+	}
+	selfIdentity, selfKnown = id, true
+	// Only an unprivileged daemon delegates its authority to its own identity.
+	// When it is root or LocalSystem, sharing its identity does not mean sharing
+	// its power: on Windows a filtered and a full token carry the same SID, so
+	// matching there would let a non-elevated shell of an administrator account
+	// act as an administrator, which is the boundary the token check exists to
+	// keep.
+	selfMayDelegate = !id.IsPrivileged()
+}
+
+// IsDaemonSelf reports whether an identity is this very process. The JSON gateway
+// runs inside the daemon and re-dials it locally, so this is what distinguishes
+// the gateway from any other caller, whatever user the daemon runs as.
+func IsDaemonSelf(id Identity) bool {
+	if !selfKnown || id.IsWindows() != selfIdentity.IsWindows() {
+		return false
+	}
+	if id.IsWindows() {
+		return id.SID != "" && id.SID == selfIdentity.SID
+	}
+	return id.UID == selfIdentity.UID
+}
+
+// IsPrivilegedCaller reports whether an identity may make the changes the daemon
+// restricts to the platform administrator. This is the daemon's own rule and
+// cannot be evaluated by a client, which does not know what the daemon runs as.
+//
+// Beyond root/administrator it accepts a caller running as the daemon's own
+// identity when the daemon is itself unprivileged. That keeps a rootless container
+// working, where there is no uid 0 at all, and a Windows daemon in netstack mode,
+// which needs no administrator rights. In those setups a caller sharing the
+// daemon's identity can already rewrite the config files it reads and replace the
+// binary it runs, so refusing it a config change would protect nothing; and an
+// unprivileged daemon cannot hand out a root shell in the first place.
+func IsPrivilegedCaller(id Identity) bool {
+	if id.IsPrivileged() {
+		return true
+	}
+	return selfMayDelegate && IsDaemonSelf(id)
+}
+
+// SelfDelegatesTo returns the identity this process delegates its authority to,
+// and whether it delegates at all. Only an unprivileged daemon does: see
+// IsPrivilegedCaller. It exists so a refusal can name who may actually perform the
+// operation, because on such a host root is neither required nor necessarily
+// available.
+func SelfDelegatesTo() (Identity, bool) {
+	if !selfKnown || !selfMayDelegate {
+		return Identity{}, false
+	}
+	return selfIdentity, true
+}
+
+// PrivilegedActor names the principal a privileged operation requires, for use
+// in messages shown to the user.
+func PrivilegedActor() string {
+	if runtime.GOOS == "windows" {
+		return "administrator privileges"
+	}
+	return "root"
+}
+
+// ElevatedCommand renders a command so that running it grants the privileges the
+// operation needs. Windows has no in-line equivalent of sudo, so the command is
+// returned unchanged and the user is expected to run it from an elevated
+// terminal.
+func ElevatedCommand(command string) string {
+	if runtime.GOOS == "windows" {
+		return command
+	}
+	return "sudo " + command
+}
+
+// UpCommand renders an elevated `netbird up` with the given flags, preceded by a
+// `down`. The down is what makes the command work on a connected client: `netbird
+// up` prints "Already connected" and returns without applying any config flag, so
+// on its own the command would appear to do nothing. It is a no-op, exit 0, when
+// the client is not connected.
+//
+// ";" rather than "&&" so the line can be pasted into any of the shells a user
+// might have: PowerShell 5.1, still the default on Windows Server, rejects "&&"
+// as a syntax error.
+func UpCommand(flags string) string {
+	return ElevatedCommand("netbird down") + "; " + ElevatedCommand("netbird up "+flags)
+}

client/internal/ipcauth/privileged_test.go

@@ -0,0 +1,134 @@
+package ipcauth
+
+import "testing"
+
+// The self rule is the one place privilege is granted to something other than the
+// platform administrator, so its two guards matter: it must apply only when the
+// daemon is itself unprivileged, and only to a caller with the daemon's identity.
+func TestIsPrivilegedCaller_SelfRule(t *testing.T) {
+	tests := []struct {
+		name string
+		// self stands in for the process the daemon runs as.
+		self      Identity
+		selfKnown bool
+		caller    Identity
+		want      bool
+	}{
+		{
+			name:      "root is privileged whatever the daemon runs as",
+			self:      Identity{UID: 1000},
+			selfKnown: true,
+			caller:    Identity{UID: 0},
+			want:      true,
+		},
+		{
+			name:      "an unprivileged daemon delegates to its own user (rootless container)",
+			self:      Identity{UID: 1000},
+			selfKnown: true,
+			caller:    Identity{UID: 1000},
+			want:      true,
+		},
+		{
+			name:      "an unprivileged daemon delegates to nobody else",
+			self:      Identity{UID: 1000},
+			selfKnown: true,
+			caller:    Identity{UID: 1001},
+			want:      false,
+		},
+		{
+			// The daemon is root on a normal install, so sharing its identity is
+			// already covered by being root; nothing else may match.
+			name:      "a root daemon delegates to nobody",
+			self:      Identity{UID: 0},
+			selfKnown: true,
+			caller:    Identity{UID: 1000},
+			want:      false,
+		},
+		{
+			// Windows netstack mode: the daemon needs no administrator rights.
+			name:      "an unprivileged windows daemon delegates to its own SID",
+			self:      Identity{SID: "S-1-5-21-1-2-3-1001"},
+			selfKnown: true,
+			caller:    Identity{SID: "S-1-5-21-1-2-3-1001"},
+			want:      true,
+		},
+		{
+			name:      "an unprivileged windows daemon delegates to no other SID",
+			self:      Identity{SID: "S-1-5-21-1-2-3-1001"},
+			selfKnown: true,
+			caller:    Identity{SID: "S-1-5-21-1-2-3-1002"},
+			want:      false,
+		},
+		{
+			// The UAC boundary: a filtered and a full token of the same account
+			// carry the same SID but not the same power, so an elevated daemon must
+			// never delegate to its own SID.
+			name:      "an elevated windows daemon does not delegate to its own SID",
+			self:      Identity{SID: "S-1-5-21-1-2-3-500", Elevated: true},
+			selfKnown: true,
+			caller:    Identity{SID: "S-1-5-21-1-2-3-500"},
+			want:      false,
+		},
+		{
+			name:      "LocalSystem is privileged on its own merits, not by delegation",
+			self:      Identity{SID: sidLocalSystem},
+			selfKnown: true,
+			caller:    Identity{SID: sidLocalSystem},
+			want:      true, // LocalSystem is privileged on its own merits
+		},
+		{
+			name:      "identities of different kinds never match",
+			self:      Identity{UID: 1000},
+			selfKnown: true,
+			caller:    Identity{SID: "S-1-5-21-1-2-3-1001"},
+			want:      false,
+		},
+		{
+			name:      "an unknown self identity delegates to nobody",
+			self:      Identity{},
+			selfKnown: false,
+			caller:    Identity{UID: 1000},
+			want:      false,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			prevID, prevKnown, prevDelegate := selfIdentity, selfKnown, selfMayDelegate
+			t.Cleanup(func() { selfIdentity, selfKnown, selfMayDelegate = prevID, prevKnown, prevDelegate })
+
+			selfIdentity, selfKnown = tt.self, tt.selfKnown
+			selfMayDelegate = tt.selfKnown && !tt.self.IsPrivileged()
+
+			if got := IsPrivilegedCaller(tt.caller); got != tt.want {
+				t.Fatalf("IsPrivilegedCaller(%v) with daemon %v = %t, want %t",
+					tt.caller, tt.self, got, tt.want)
+			}
+		})
+	}
+}
+
+// The real process must never accidentally delegate: a test binary running as a
+// normal user is unprivileged, so it may match itself, but nothing else.
+func TestIsPrivilegedCaller_ThisProcess(t *testing.T) {
+	id, err := CurrentProcessIdentity()
+	if err != nil {
+		t.Skipf("cannot read this process's identity: %v", err)
+	}
+
+	// This process is always allowed to act as itself: either it is privileged, or
+	// it is unprivileged and therefore delegates to its own identity.
+	if !IsPrivilegedCaller(id) {
+		t.Errorf("this process %v was refused its own identity", id)
+	}
+
+	// A caller that is neither root nor this process must be refused, whatever
+	// this process happens to be.
+	other := Identity{UID: id.UID + 1}
+	if id.IsWindows() {
+		other = Identity{SID: id.SID + "9"}
+	}
+	if IsPrivilegedCaller(other) {
+		t.Errorf("an unrelated identity %v was treated as privileged", other)
+	}
+}

client/internal/ipcauth/self_unix.go

@@ -0,0 +1,17 @@
+//go:build !windows
+
+package ipcauth
+
+import "os"
+
+// CurrentProcessIdentity returns this process's identity as the daemon would
+// see it if this process connected to the local IPC. It lets a client (the UI)
+// decide up front whether a privileged operation can succeed, without a
+// round-trip and without duplicating the rules: the answer comes from the same
+// Identity.IsPrivileged the daemon applies.
+func CurrentProcessIdentity() (Identity, error) {
+	return Identity{
+		UID: uint32(os.Geteuid()),
+		GID: uint32(os.Getegid()),
+	}, nil
+}

client/internal/ipcauth/self_windows.go

@@ -0,0 +1,35 @@
+//go:build windows
+
+package ipcauth
+
+import (
+	"fmt"
+
+	"golang.org/x/sys/windows"
+)
+
+// CurrentProcessIdentity returns this process's identity as the daemon would see
+// it if this process connected to the local IPC. It lets a client (the UI)
+// decide up front whether a privileged operation can succeed, without a
+// round-trip and without duplicating the rules: the answer comes from the same
+// Identity.IsPrivileged the daemon applies to the token it reads off the pipe.
+func CurrentProcessIdentity() (Identity, error) {
+	// A pseudo-token, so it must not be closed.
+	token := windows.GetCurrentProcessToken()
+
+	user, err := token.GetTokenUser()
+	if err != nil {
+		return Identity{}, fmt.Errorf("read token user: %w", err)
+	}
+
+	groups, err := tokenGroupSIDs(token)
+	if err != nil {
+		return Identity{}, err
+	}
+
+	return Identity{
+		SID:      user.User.Sid.String(),
+		Groups:   groups,
+		Elevated: token.IsElevated(),
+	}, nil
+}

client/internal/profilemanager/config.go

@@ -746,6 +746,13 @@ func (config *Config) applyMDMPolicy(policy *mdm.Policy) {
 // appended for https or ":80" for http. The serviceName parameter is
 // used to contextualise error messages. On success returns the parsed
 // *url.URL; on failure returns a non-nil error.
+// ParseServiceURL normalises a service URL exactly as the config layer does when
+// it stores one, so callers comparing a requested URL against a stored one do not
+// have to reimplement the scheme validation and default-port handling.
+func ParseServiceURL(serviceName, serviceURL string) (*url.URL, error) {
+	return parseURL(serviceName, serviceURL)
+}
+
 func parseURL(serviceName, serviceURL string) (*url.URL, error) {
 	parsedMgmtURL, err := url.ParseRequestURI(serviceURL)
 	if err != nil {

client/server/login_gate_test.go

@@ -0,0 +1,127 @@
+package server
+
+import (
+	"context"
+	"path/filepath"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+	"google.golang.org/grpc/codes"
+	gstatus "google.golang.org/grpc/status"
+
+	"github.com/netbirdio/netbird/client/internal"
+	"github.com/netbirdio/netbird/client/internal/profilemanager"
+	"github.com/netbirdio/netbird/client/proto"
+)
+
+// A refused login must not leave the profile switched. Login can both switch
+// profiles and carry the guarded config fields, so the gate has to run before the
+// switch: otherwise a caller whose change is refused still gets the side effect of
+// activating whichever profile the request named.
+func TestLogin_RefusedChangeLeavesTheProfileAlone(t *testing.T) {
+	s, _, activeProfile, username, _ := setupServerWithProfile(t)
+
+	// Login reads process state off the daemon's root context.
+	s.rootCtx = internal.CtxInitState(context.Background())
+
+	// A second profile that runs the SSH server, which is what makes repointing
+	// its management binding a privileged change.
+	target := "ssh-enabled"
+	_, err := profilemanager.UpdateOrCreateConfig(profilemanager.ConfigInput{
+		ConfigPath:       filepath.Join(profilemanager.DefaultConfigPathDir, target+".json"),
+		ManagementURL:    "https://api.netbird.io:443",
+		ServerSSHAllowed: boolPtr(true),
+	})
+	require.NoError(t, err)
+
+	_, err = s.Login(userCtx(), &proto.LoginRequest{
+		ProfileName:   &target,
+		Username:      &username,
+		ManagementUrl: "https://mgmt.attacker.example:443",
+	})
+	require.Error(t, err, "an unprivileged caller must not move the management URL of an SSH-enabled profile")
+	require.Equal(t, codes.PermissionDenied, gstatus.Code(err), "want a privilege refusal, got %v", err)
+
+	active, err := s.profileManager.GetActiveProfileState()
+	require.NoError(t, err)
+	require.Equal(t, profilemanager.ID(activeProfile), active.ID,
+		"the refused login switched the active profile anyway")
+}
+
+// A caller whose change becomes privileged only after its first check must be
+// refused without having cancelled a login or switched profiles: the first check is
+// unsynchronized, so the SSH server can be enabled by a concurrent privileged
+// request in between, and the authoritative check happens before any side effect.
+func TestLogin_ChangeThatBecomesPrivilegedMidRequestHasNoSideEffects(t *testing.T) {
+	s, _, activeProfile, username, _ := setupServerWithProfile(t)
+	s.rootCtx = internal.CtxInitState(context.Background())
+
+	// The target profile has SSH off, so the first check lets the request through.
+	target := "ssh-later"
+	targetPath := filepath.Join(profilemanager.DefaultConfigPathDir, target+".json")
+	_, err := profilemanager.UpdateOrCreateConfig(profilemanager.ConfigInput{
+		ConfigPath:       targetPath,
+		ManagementURL:    "https://api.netbird.io:443",
+		ServerSSHAllowed: boolPtr(false),
+	})
+	require.NoError(t, err)
+
+	cancelled := false
+	s.actCancel = func() { cancelled = true }
+
+	// Stand in for a privileged SetConfig that enables the SSH server between the
+	// two checks, which is the interleaving the lock has to make safe.
+	afterLoginPreCheck = func() {
+		_, err := profilemanager.UpdateOrCreateConfig(profilemanager.ConfigInput{
+			ConfigPath:       targetPath,
+			ServerSSHAllowed: boolPtr(true),
+		})
+		require.NoError(t, err)
+	}
+	t.Cleanup(func() { afterLoginPreCheck = nil })
+
+	_, err = s.Login(userCtx(), &proto.LoginRequest{
+		ProfileName:   &target,
+		Username:      &username,
+		ManagementUrl: "https://mgmt.attacker.example:443",
+	})
+	require.Error(t, err)
+	require.Equal(t, codes.PermissionDenied, gstatus.Code(err), "want a privilege refusal, got %v", err)
+	require.False(t, cancelled, "the refused login cancelled the login already in progress")
+
+	active, err := s.profileManager.GetActiveProfileState()
+	require.NoError(t, err)
+	require.Equal(t, profilemanager.ID(activeProfile), active.ID, "the refused login switched the active profile anyway")
+
+	stored, err := profilemanager.ReadConfig(targetPath)
+	require.NoError(t, err)
+	require.Equal(t, "https://api.netbird.io:443", stored.ManagementURL.String(), "the refused login moved the management URL")
+}
+
+// Login cancels whatever login is already in progress before starting its own. A
+// refused caller must not get that far, otherwise anyone able to reach the socket
+// can abort someone else's login by sending a request that is denied.
+func TestLogin_RefusedChangeLeavesAnInProgressLoginAlone(t *testing.T) {
+	s, _, _, username, _ := setupServerWithProfile(t)
+	s.rootCtx = internal.CtxInitState(context.Background())
+
+	target := "ssh-enabled"
+	_, err := profilemanager.UpdateOrCreateConfig(profilemanager.ConfigInput{
+		ConfigPath:       filepath.Join(profilemanager.DefaultConfigPathDir, target+".json"),
+		ManagementURL:    "https://api.netbird.io:443",
+		ServerSSHAllowed: boolPtr(true),
+	})
+	require.NoError(t, err)
+
+	cancelled := false
+	s.actCancel = func() { cancelled = true }
+
+	_, err = s.Login(userCtx(), &proto.LoginRequest{
+		ProfileName:   &target,
+		Username:      &username,
+		ManagementUrl: "https://mgmt.attacker.example:443",
+	})
+	require.Error(t, err)
+	require.Equal(t, codes.PermissionDenied, gstatus.Code(err), "want a privilege refusal, got %v", err)
+	require.False(t, cancelled, "the refused login cancelled the login already in progress")
+}

client/server/server.go

@@ -82,6 +82,12 @@ type Server struct {
 	// extend flow or vice versa.
 	extendAuthSessionFlow *auth.PendingFlow
 
+	// guardedConfigMu serializes a privilege check against the write it
+	// authorizes. Without it the two are separate steps over the same file, and a
+	// change that was allowed because the profile had the SSH server disabled
+	// could land after a concurrent privileged request enabled it.
+	guardedConfigMu sync.Mutex
+
 	mutex  sync.Mutex
 	config *profilemanager.Config
 	proto.UnimplementedDaemonServiceServer
@@ -411,6 +417,20 @@ func (s *Server) SetConfig(callerCtx context.Context, msg *proto.SetConfigReques
 		return nil, err
 	}
 
+	// Privilege gate: refuse the parts of the request that would let a local
+	// user turn the root daemon into a root shell. Held across the write so the
+	// config cannot gain the SSH server between the decision and the update.
+	s.guardedConfigMu.Lock()
+	defer s.guardedConfigMu.Unlock()
+
+	stored, err := s.storedProfileConfig(msg.ProfileName, msg.Username)
+	if err != nil {
+		return nil, err
+	}
+	if err := requirePrivilegeForConfigChange(callerCtx, stored, privilegedChangeFromSetConfig(msg)); err != nil {
+		return nil, err
+	}
+
 	config, err := s.setConfigInputFromRequest(msg)
 	if err != nil {
 		return nil, err
@@ -537,22 +557,23 @@ func (s *Server) Login(callerCtx context.Context, msg *proto.LoginRequest) (*pro
 		}
 	}
 
-	s.mutex.Lock()
-	if s.actCancel != nil {
-		s.actCancel()
-	}
-	ctx, cancel := context.WithCancel(callerCtx)
-
-	md, ok := metadata.FromIncomingContext(callerCtx)
-	if ok {
-		ctx = metadata.NewOutgoingContext(ctx, md)
+	activeProf, err := s.profileManager.GetActiveProfileState()
+	if err != nil {
+		log.Errorf("failed to get active profile state: %v", err)
+		return nil, fmt.Errorf("failed to get active profile state: %w", err)
 	}
 
-	s.actCancel = cancel
-	s.mutex.Unlock()
-
-	if err := RestoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
-		log.Warnf(errRestoreResidualState, err)
+	// Privilege gate: same restrictions as SetConfig, since LoginRequest can carry
+	// the same fields. It runs before anything here changes daemon state, so a
+	// refused login neither switches the profile nor cancels a login already in
+	// progress, and it reads the profile the request targets, which is the one the
+	// switch below would activate.
+	stored, err := s.storedLoginConfig(activeProf, msg)
+	if err != nil {
+		return nil, err
+	}
+	if err := requirePrivilegeForConfigChange(callerCtx, stored, privilegedChangeFromLogin(msg)); err != nil {
+		return nil, err
 	}
 
 	state := internal.CtxGetState(s.rootCtx)
@@ -563,23 +584,16 @@ func (s *Server) Login(callerCtx context.Context, msg *proto.LoginRequest) (*pro
 		}
 	}()
 
-	activeProf, err := s.profileManager.GetActiveProfileState()
+	ctx, activeProf, err := s.authorizeAndPrepareLogin(callerCtx, msg, activeProf)
 	if err != nil {
-		log.Errorf("failed to get active profile state: %v", err)
-		return nil, fmt.Errorf("failed to get active profile state: %w", err)
-	}
-
-	if msg.ProfileName != nil {
-		if _, err := s.switchProfileIfNeeded(*msg.ProfileName, msg.Username, activeProf); err != nil {
-			log.Errorf("failed to switch profile: %v", err)
-			return nil, err
+		// The RPC boundary is where this gets recorded: nothing logs handler
+		// errors for us, and a caller that retries would otherwise leave no
+		// trace in the daemon log. A refusal is skipped because the gate has
+		// already logged the decision, with the caller's identity.
+		if gstatus.Code(err) != codes.PermissionDenied {
+			log.Errorf("failed to prepare login: %v", err)
 		}
-	}
-
-	activeProf, err = s.profileManager.GetActiveProfileState()
-	if err != nil {
-		log.Errorf("failed to get active profile state: %v", err)
-		return nil, fmt.Errorf("failed to get active profile state: %w", err)
+		return nil, err
 	}
 
 	log.Infof("active profile: %s for %s", activeProf.ID, activeProf.Username)
@@ -593,11 +607,6 @@ func (s *Server) Login(callerCtx context.Context, msg *proto.LoginRequest) (*pro
 
 	s.mutex.Unlock()
 
-	if err := persistLoginOverrides(activeProf, msg.ManagementUrl, msg.OptionalPreSharedKey); err != nil {
-		log.Errorf("failed to persist login overrides: %v", err)
-		return nil, fmt.Errorf("persist login overrides: %w", err)
-	}
-
 	config, _, err := s.getConfig(activeProf)
 	if err != nil {
 		log.Errorf("failed to get active profile config: %v", err)
@@ -980,6 +989,63 @@ func (s *Server) waitForUp(callerCtx context.Context) (*proto.UpResponse, error)
 	}
 }
 
+// storedProfileConfig loads the on-disk config of the profile a request
+// targets, so a privileged-change decision can be made against the values the
+// profile currently holds. A profile that has no config file yet yields nil,
+// which every caller must read as "nothing enabled yet".
+func (s *Server) storedProfileConfig(handle, username string) (*profilemanager.Config, error) {
+	resolved, err := s.resolveProfileHandle(handle, username)
+	if err != nil {
+		return nil, err
+	}
+
+	path := resolved.Path
+	if path == "" {
+		path = profilemanager.DefaultConfigPath
+	}
+
+	return s.storedConfigAtPath(path)
+}
+
+// storedLoginConfig loads the on-disk config of the profile a login request
+// targets: the one it names, or the active one when it names none. Used to decide
+// a privileged change before the request is allowed to switch profiles.
+func (s *Server) storedLoginConfig(activeProf *profilemanager.ActiveProfileState, msg *proto.LoginRequest) (*profilemanager.Config, error) {
+	if msg.ProfileName == nil {
+		cfgPath, err := activeProf.FilePath()
+		if err != nil {
+			return nil, fmt.Errorf("active profile file path: %w", err)
+		}
+		return s.storedConfigAtPath(cfgPath)
+	}
+
+	// Mirrors switchProfileIfNeeded: the default profile resolves without a
+	// username, so this reads the same profile the switch would activate.
+	handle := *msg.ProfileName
+	username := ""
+	if handle != profilemanager.DefaultProfileName {
+		username = msg.GetUsername()
+	}
+	return s.storedProfileConfig(handle, username)
+}
+
+// storedConfigAtPath reads a profile config file, yielding nil when it does not
+// exist yet.
+func (s *Server) storedConfigAtPath(path string) (*profilemanager.Config, error) {
+	if _, err := os.Stat(path); err != nil {
+		if os.IsNotExist(err) {
+			return nil, nil //nolint:nilnil
+		}
+		return nil, fmt.Errorf("stat profile config: %w", err)
+	}
+
+	cfg, err := profilemanager.GetConfig(path)
+	if err != nil {
+		return nil, fmt.Errorf("read profile config: %w", err)
+	}
+	return cfg, nil
+}
+
 // resolveProfileHandle resolves a wire-level profile handle (display
 // name, ID, or unique ID prefix) to a concrete profile. Returns gRPC
 // status errors so handlers can return them directly.
@@ -1197,6 +1263,12 @@ func (s *Server) handleProfileLogout(ctx context.Context, msg *proto.LogoutReque
 
 	if err := s.logoutFromProfile(ctx, resolved); err != nil {
 		log.Errorf("failed to logout from profile %s: %v", resolved.ID, err)
+		// A refused deregistration is already a status error carrying the reason
+		// and the command to run; rewrapping it as Internal would flatten both
+		// into a gRPC dump for the user.
+		if _, isStatus := gstatus.FromError(err); isStatus {
+			return nil, err
+		}
 		return nil, gstatus.Errorf(codes.Internal, "logout: %v", err)
 	}
 
@@ -1318,6 +1390,13 @@ func (s *Server) sendLogoutRequest(ctx context.Context) error {
 }
 
 func (s *Server) sendLogoutRequestWithConfig(ctx context.Context, config *profilemanager.Config) error {
+	// Privilege gate: deregistering frees this machine's key to be registered
+	// against another management server, which is only restricted while the SSH
+	// server makes that a privilege handover.
+	if err := requirePrivilegeForDeregistration(ctx, config); err != nil {
+		return err
+	}
+
 	key, err := wgtypes.ParseKey(config.PrivateKey)
 	if err != nil {
 		return fmt.Errorf("parse private key: %w", err)
@@ -2063,7 +2142,10 @@ func (s *Server) RemoveProfile(ctx context.Context, msg *proto.RemoveProfileRequ
 	}
 
 	if err := s.logoutFromProfile(ctx, resolved); err != nil {
-		log.Warnf("failed to logout from profile %s before removal: %v", resolved.ID, err)
+		// Deregistration is best-effort here: the local profile is removed
+		// either way, so an unprivileged caller leaves the peer registered on
+		// the management server rather than being blocked from removing it.
+		log.Warnf("removing profile %s locally without deregistering it: %v", resolved.ID, err)
 	}
 
 	if err := s.profileManager.RemoveProfile(resolved.ID, msg.Username); err != nil {
@@ -2360,6 +2442,69 @@ func sendTerminalNotification() error {
 
 // persistLoginOverrides writes management URL and pre-shared key from a LoginRequest to the
 // active profile config so that subsequent reads pick them up. Empty/nil values are ignored.
+// afterLoginPreCheck is a seam for tests to run a concurrent config change
+// between Login's first privilege check and the authoritative one.
+var afterLoginPreCheck func()
+
+// authorizeAndPrepareLogin makes the authoritative privilege decision for a login
+// and, when it passes, carries out every state change that decision authorizes:
+// cancelling an login already in progress, switching to the requested profile, and
+// persisting the config overrides the request carries.
+//
+// All of it happens under guardedConfigMu, which SetConfig also holds across its
+// own check and write. Login's earlier check refuses the ordinary case before any
+// of this is reached; this one exists because that check is not synchronized
+// against a concurrent privileged request that enables the SSH server, and a
+// caller refused here must not have cancelled or switched anything either.
+func (s *Server) authorizeAndPrepareLogin(callerCtx context.Context, msg *proto.LoginRequest, activeProf *profilemanager.ActiveProfileState) (context.Context, *profilemanager.ActiveProfileState, error) {
+	if afterLoginPreCheck != nil {
+		afterLoginPreCheck()
+	}
+
+	s.guardedConfigMu.Lock()
+	defer s.guardedConfigMu.Unlock()
+
+	stored, err := s.storedLoginConfig(activeProf, msg)
+	if err != nil {
+		return nil, nil, err
+	}
+	if err := requirePrivilegeForConfigChange(callerCtx, stored, privilegedChangeFromLogin(msg)); err != nil {
+		return nil, nil, err
+	}
+
+	s.mutex.Lock()
+	if s.actCancel != nil {
+		s.actCancel()
+	}
+	ctx, cancel := context.WithCancel(callerCtx)
+	if md, ok := metadata.FromIncomingContext(callerCtx); ok {
+		ctx = metadata.NewOutgoingContext(ctx, md)
+	}
+	s.actCancel = cancel
+	s.mutex.Unlock()
+
+	if err := RestoreResidualState(s.rootCtx, s.profileManager.GetStatePath()); err != nil {
+		log.Warnf(errRestoreResidualState, err)
+	}
+
+	if msg.ProfileName != nil {
+		if _, err := s.switchProfileIfNeeded(*msg.ProfileName, msg.Username, activeProf); err != nil {
+			return nil, nil, fmt.Errorf("switch profile: %w", err)
+		}
+	}
+
+	activeProf, err = s.profileManager.GetActiveProfileState()
+	if err != nil {
+		return nil, nil, fmt.Errorf("active profile state: %w", err)
+	}
+
+	if err := persistLoginOverrides(activeProf, msg.ManagementUrl, msg.OptionalPreSharedKey); err != nil {
+		return nil, nil, fmt.Errorf("persist login overrides: %w", err)
+	}
+
+	return ctx, activeProf, nil
+}
+
 func persistLoginOverrides(activeProf *profilemanager.ActiveProfileState, managementURL string, preSharedKey *string) error {
 	if preSharedKey != nil && *preSharedKey == "" {
 		preSharedKey = nil

client/server/setconfig_mdm_test.go

@@ -66,7 +66,11 @@ func setupServerWithProfile(t *testing.T) (s *Server, ctx context.Context, profN
 		Username: currUser.Username,
 	}))
 
-	ctx = context.Background()
+	// The privileged-change gate reads the caller's kernel identity from the
+	// context, which a real caller gets from the daemon's transport credentials.
+	// This test drives the handler directly, so it stands in for a root caller;
+	// without an identity the gate would (correctly) refuse the SSH fields.
+	ctx = privilegedTestCtx()
 	s = New(ctx, "console", "", false, false, false, false)
 	return s, ctx, profName, currUser.Username, cfgPath
 }

client/server/setconfig_test.go

@@ -1,7 +1,6 @@
 package server
 
 import (
-	"context"
 	"os/user"
 	"path/filepath"
 	"reflect"
@@ -52,7 +51,11 @@ func TestSetConfig_AllFieldsSaved(t *testing.T) {
 	})
 	require.NoError(t, err)
 
-	ctx := context.Background()
+	// The privileged-change gate reads the caller's kernel identity from the
+	// context, which a real caller gets from the daemon's transport credentials.
+	// This test drives the handler directly, so it stands in for a root caller;
+	// without an identity the gate would (correctly) refuse the SSH fields.
+	ctx := privilegedTestCtx()
 	s := New(ctx, "console", "", false, false, false, false)
 
 	rosenpassEnabled := true

client/server/ssh_gate.go

@@ -0,0 +1,282 @@
+package server
+
+import (
+	"context"
+	"fmt"
+	"net/url"
+	"runtime"
+	"strings"
+
+	log "github.com/sirupsen/logrus"
+	"google.golang.org/genproto/googleapis/rpc/errdetails"
+	"google.golang.org/grpc/codes"
+	gstatus "google.golang.org/grpc/status"
+
+	"github.com/netbirdio/netbird/client/internal/daemonaddr"
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
+	"github.com/netbirdio/netbird/client/internal/profilemanager"
+	"github.com/netbirdio/netbird/client/proto"
+	"github.com/netbirdio/netbird/util"
+)
+
+// The daemon runs as root/LocalSystem, so a handful of config changes cross the
+// user-to-root boundary and are restricted to privileged callers:
+//
+//   - Enabling SSH root login, or disabling SSH authentication, turns the
+//     daemon's SSH server into a root (or unauthenticated) shell.
+//   - Enabling the SSH server at all is what makes the above reachable, and a
+//     profile the caller owns is not a privilege they hold.
+//   - While the SSH server is enabled, repointing the profile at another
+//     management identity hands SSH authorization decisions, including which
+//     keys and users are accepted, to whoever controls that identity. Changing
+//     the management URL and deregistering the peer are both ways to do that.
+//
+// Everything else stays unauthenticated, so this is not an authorization model:
+// it only refuses the changes that would let a local user become root. A caller
+// whose identity cannot be established is refused as well.
+
+// privilegedConfigChange is the subset of a config request that crosses the
+// user-to-root boundary. Fields are nil or empty when the request leaves them
+// untouched.
+type privilegedConfigChange struct {
+	managementURL    string
+	serverSSHAllowed *bool
+	enableSSHRoot    *bool
+	disableSSHAuth   *bool
+}
+
+func privilegedChangeFromSetConfig(msg *proto.SetConfigRequest) privilegedConfigChange {
+	return privilegedConfigChange{
+		managementURL:    msg.GetManagementUrl(),
+		serverSSHAllowed: msg.ServerSSHAllowed,
+		enableSSHRoot:    msg.EnableSSHRoot,
+		disableSSHAuth:   msg.DisableSSHAuth,
+	}
+}
+
+func privilegedChangeFromLogin(msg *proto.LoginRequest) privilegedConfigChange {
+	return privilegedConfigChange{
+		managementURL:    msg.GetManagementUrl(),
+		serverSSHAllowed: msg.ServerSSHAllowed,
+		enableSSHRoot:    msg.EnableSSHRoot,
+		disableSSHAuth:   msg.DisableSSHAuth,
+	}
+}
+
+// requirePrivilegeForConfigChange refuses the privileged parts of a config
+// change when the caller is not root/administrator. stored is the profile's
+// current config, or nil when it has none yet.
+//
+// Each check compares against the stored value so that a request restating a
+// value it does not change is never refused: a UI that submits the whole
+// settings form must not start failing once an administrator has enabled SSH.
+func requirePrivilegeForConfigChange(ctx context.Context, stored *profilemanager.Config, change privilegedConfigChange) error {
+	if enables(storedFlag(stored, func(c *profilemanager.Config) *bool { return c.EnableSSHRoot }), change.enableSSHRoot) {
+		return denyPrivileged(ctx, "enabling SSH root login", ipcauth.UpCommand("--enable-ssh-root"))
+	}
+
+	if enables(storedFlag(stored, func(c *profilemanager.Config) *bool { return c.DisableSSHAuth }), change.disableSSHAuth) {
+		return denyPrivileged(ctx, "disabling SSH authentication", ipcauth.UpCommand("--disable-ssh-auth"))
+	}
+
+	if enables(sshServerCurrentlyAllowed(stored), change.serverSSHAllowed) {
+		return denyPrivileged(ctx, "enabling the NetBird SSH server", ipcauth.UpCommand("--allow-server-ssh"))
+	}
+
+	// Only guard the management binding while the SSH server is enabled: that is
+	// when the management identity decides who may open a shell here.
+	if !sshServerEnabled(stored) {
+		return nil
+	}
+
+	if change.managementURL != "" && !sameManagementURL(stored.ManagementURL, change.managementURL) {
+		return denyPrivileged(ctx,
+			"changing the management URL while the NetBird SSH server is enabled",
+			ipcauth.UpCommand("-m "+change.managementURL))
+	}
+
+	return nil
+}
+
+// requirePrivilegeForDeregistration refuses to deregister the peer from the
+// management server when the caller is not privileged and the profile has the
+// SSH server enabled. Deregistering frees the peer's key to be registered
+// against another management identity, which is the same handover the
+// management URL check refuses.
+//
+// Callers that treat deregistration as best-effort (profile removal) continue
+// without it; callers that were asked to deregister surface the error.
+func requirePrivilegeForDeregistration(ctx context.Context, cfg *profilemanager.Config) error {
+	if !sshServerEnabled(cfg) {
+		return nil
+	}
+
+	return denyPrivileged(ctx,
+		"deregistering this peer while the NetBird SSH server is enabled",
+		ipcauth.ElevatedCommand("netbird logout"))
+}
+
+// denyPrivileged returns nil when the caller is privileged, and otherwise a
+// PermissionDenied whose message names the action and the command that performs
+// it with the privileges it needs. The same summary and command ride along as an
+// ErrorInfo detail so the CLI and the UI can present them without parsing text.
+//
+// action reads as the subject of a sentence ("enabling SSH root login"), and
+// command is the equivalent command, already elevated for the platform.
+func denyPrivileged(ctx context.Context, action, command string) error {
+	id, ok := ipcauth.CallerIdentity(ctx)
+	if !ok {
+		log.Warnf("denying %s: the caller's identity cannot be verified on this control channel", action)
+		return privilegeError(unidentifiedSummary(action), reinstallCommand())
+	}
+
+	if ipcauth.IsPrivilegedCaller(id) {
+		log.Infof("allowing %s for privileged caller %s", action, id)
+		return nil
+	}
+
+	log.Warnf("denying %s for unprivileged caller %s", action, id)
+	actor, command := requiredActor(command)
+	return privilegeError(privilegeSummary(action, actor), command)
+}
+
+// requiredActor names who may perform the operation and adjusts the command to
+// match. A daemon that is not itself privileged delegates to its own identity, so
+// telling that host's user to become root is wrong twice over: root is not what the
+// daemon checks for, and a rootless container has neither root nor sudo.
+func requiredActor(command string) (string, string) {
+	self, delegates := ipcauth.SelfDelegatesTo()
+	if !delegates {
+		return ipcauth.PrivilegedActor(), command
+	}
+	return fmt.Sprintf("the user the daemon runs as (%s)", self), strings.ReplaceAll(command, "sudo ", "")
+}
+
+// privilegeError builds the PermissionDenied carrying summary and command.
+func privilegeError(summary, command string) error {
+	st := gstatus.New(codes.PermissionDenied, fmt.Sprintf("%s\n\n%s", summary, command))
+
+	detailed, err := st.WithDetails(&errdetails.ErrorInfo{
+		Reason: ipcauth.ErrorReasonPrivilegeRequired,
+		Domain: ipcauth.ErrorDomain,
+		Metadata: map[string]string{
+			ipcauth.ErrorMetaSummary: summary,
+			ipcauth.ErrorMetaCommand: command,
+		},
+	})
+	if err != nil {
+		log.Debugf("attach privilege error detail: %v", err)
+		return st.Err()
+	}
+	return detailed.Err()
+}
+
+// privilegeSummary states what is refused and what it needs, in one sentence
+// that reads the same in a dialog and in a terminal.
+func privilegeSummary(action, actor string) string {
+	return fmt.Sprintf("%s requires %s.", capitalize(action), actor)
+}
+
+// unidentifiedSummary covers a control channel that carries no caller identity.
+// Elevating does not help there, so it points at the daemon's socket instead.
+func unidentifiedSummary(action string) string {
+	return fmt.Sprintf("%s requires %s, and the daemon cannot verify who is calling over its current socket. "+
+		"Reinstall the service on a socket that carries the caller's identity.", capitalize(action), ipcauth.PrivilegedActor())
+}
+
+// reinstallCommand is the command that moves the daemon onto a socket whose
+// callers can be identified.
+func reinstallCommand() string {
+	if runtime.GOOS == "windows" {
+		return fmt.Sprintf("netbird service install --daemon-addr %s", daemonaddr.WindowsPipeAddr)
+	}
+	return "sudo netbird service install --daemon-addr unix:///var/run/netbird.sock"
+}
+
+func capitalize(s string) string {
+	if s == "" {
+		return s
+	}
+	return strings.ToUpper(s[:1]) + s[1:]
+}
+
+// enables reports whether requested turns a flag on that is currently off. A
+// request that restates the stored value, or turns the flag off, is not a
+// privileged change.
+func enables(stored, requested *bool) bool {
+	if requested == nil || !*requested {
+		return false
+	}
+	return stored == nil || !*stored
+}
+
+// storedFlag reads a flag from the stored config, tolerating a config that does
+// not exist yet.
+func storedFlag(cfg *profilemanager.Config, get func(*profilemanager.Config) *bool) *bool {
+	if cfg == nil {
+		return nil
+	}
+	return get(cfg)
+}
+
+// sshServerEnabled reports whether the profile currently runs the SSH server.
+//
+// A nil flag means ON, matching what the engine does with the same config
+// (util.ReturnBoolWithDefaultTrue in internal/connect.go, kept for configs written
+// before the flag existed). Reading it as OFF here would open the management-URL
+// and deregistration guards on exactly those legacy hosts, whose SSH server is
+// running. Configs loaded through profilemanager have already been materialised by
+// apply(), so this is the same answer by a route that does not depend on that.
+func sshServerEnabled(cfg *profilemanager.Config) bool {
+	if cfg == nil {
+		return false
+	}
+	return util.ReturnBoolWithDefaultTrue(cfg.ServerSSHAllowed)
+}
+
+// sshServerCurrentlyAllowed is the value an enable request is compared against. It
+// shares sshServerEnabled's nil-means-on default, so restating "on" for a legacy
+// config is correctly seen as no change.
+func sshServerCurrentlyAllowed(cfg *profilemanager.Config) *bool {
+	enabled := sshServerEnabled(cfg)
+	if cfg == nil {
+		return nil
+	}
+	return &enabled
+}
+
+// sameManagementURL reports whether requested addresses the same management
+// server as stored, comparing scheme, host and effective port so that an
+// equivalent spelling ("https://api.netbird.io" for a stored
+// "https://api.netbird.io:443") is not treated as a change. It fails closed:
+// anything unparseable counts as a change and therefore needs privilege.
+func sameManagementURL(stored *url.URL, requested string) bool {
+	if stored == nil {
+		return false
+	}
+
+	// Normalise the requested URL through the config layer's own parser, so the
+	// comparison cannot drift from how the value would actually be stored.
+	parsed, err := profilemanager.ParseServiceURL("Management URL", requested)
+	if err != nil {
+		return false
+	}
+
+	return stored.Scheme == parsed.Scheme &&
+		stored.Hostname() == parsed.Hostname() &&
+		effectivePort(stored) == effectivePort(parsed)
+}
+
+func effectivePort(u *url.URL) string {
+	if port := u.Port(); port != "" {
+		return port
+	}
+	switch u.Scheme {
+	case "https":
+		return "443"
+	case "http":
+		return "80"
+	default:
+		return ""
+	}
+}

client/server/ssh_gate_test.go

@@ -0,0 +1,348 @@
+package server
+
+import (
+	"context"
+	"net/url"
+	"os"
+	"runtime"
+	"strings"
+	"testing"
+
+	"google.golang.org/genproto/googleapis/rpc/errdetails"
+	"google.golang.org/grpc/codes"
+	"google.golang.org/grpc/credentials"
+	"google.golang.org/grpc/peer"
+	gstatus "google.golang.org/grpc/status"
+
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
+	"github.com/netbirdio/netbird/client/internal/profilemanager"
+)
+
+// ctxWithIdentity builds a request context carrying the identity the transport
+// credentials would have attached.
+func ctxWithIdentity(id ipcauth.Identity) context.Context {
+	return peer.NewContext(context.Background(), &peer.Peer{
+		AuthInfo: ipcauth.AuthInfo{
+			CommonAuthInfo: credentials.CommonAuthInfo{SecurityLevel: credentials.NoSecurity},
+			Identity:       id,
+		},
+	})
+}
+
+// unprivUID is deliberately not this process's own uid. An unprivileged daemon
+// treats a caller sharing its identity as privileged (rootless containers), and
+// the test binary would otherwise stand in for both the daemon and the caller.
+// os.Geteuid returns -1 on Windows, where identities are SIDs instead and this is
+// unused.
+var unprivUID = uint32(os.Geteuid() + 1)
+
+// The fabricated identities have to be shaped like the platform's: a uid says
+// nothing on Windows, and a zero uid there would read as root and be privileged.
+func rootCtx() context.Context { return ctxWithIdentity(privilegedIdentity()) }
+func userCtx() context.Context { return ctxWithIdentity(unprivilegedIdentity()) }
+
+func privilegedIdentity() ipcauth.Identity {
+	if runtime.GOOS == "windows" {
+		// LocalSystem, which is what the Windows service account is.
+		return ipcauth.Identity{SID: "S-1-5-18"}
+	}
+	return ipcauth.Identity{UID: 0}
+}
+
+func unprivilegedIdentity() ipcauth.Identity {
+	if runtime.GOOS == "windows" {
+		// A plain user SID: no groups, so no BUILTIN\Administrators, and not
+		// elevated.
+		return ipcauth.Identity{SID: "S-1-5-21-1-2-3-1001"}
+	}
+	return ipcauth.Identity{UID: unprivUID, GID: unprivUID}
+}
+func noIdentityCtx() context.Context { return context.Background() }
+
+func boolPtr(v bool) *bool { return &v }
+
+func mustURL(t *testing.T, raw string) *url.URL {
+	t.Helper()
+	u, err := url.Parse(raw)
+	if err != nil {
+		t.Fatalf("parse %q: %v", raw, err)
+	}
+	return u
+}
+
+func assertDenied(t *testing.T, err error) {
+	t.Helper()
+	if err == nil {
+		t.Fatal("expected the change to be refused, got nil")
+	}
+	st := gstatus.Convert(err)
+	if st.Code() != codes.PermissionDenied {
+		t.Fatalf("code = %v, want PermissionDenied", st.Code())
+	}
+	// The refusal must be machine-readable: the CLI and the UI render the
+	// summary and command from the detail rather than parsing the message.
+	var info *errdetails.ErrorInfo
+	for _, d := range st.Details() {
+		if got, ok := d.(*errdetails.ErrorInfo); ok {
+			info = got
+		}
+	}
+	if info == nil {
+		t.Fatal("refusal carries no ErrorInfo detail")
+	}
+	if info.GetReason() != ipcauth.ErrorReasonPrivilegeRequired || info.GetDomain() != ipcauth.ErrorDomain {
+		t.Fatalf("detail = %s/%s, want %s/%s", info.GetDomain(), info.GetReason(), ipcauth.ErrorDomain, ipcauth.ErrorReasonPrivilegeRequired)
+	}
+	if info.GetMetadata()[ipcauth.ErrorMetaSummary] == "" {
+		t.Error("detail carries no summary")
+	}
+	if info.GetMetadata()[ipcauth.ErrorMetaCommand] == "" {
+		t.Error("detail carries no command")
+	}
+}
+
+func assertAllowed(t *testing.T, err error) {
+	t.Helper()
+	if err != nil {
+		t.Fatalf("expected the change to be allowed, got %v", err)
+	}
+}
+
+func TestRequirePrivilegeForConfigChange_SSHFlags(t *testing.T) {
+	tests := []struct {
+		name       string
+		stored     *profilemanager.Config
+		change     privilegedConfigChange
+		privileged bool
+		wantDeny   bool
+	}{
+		{
+			name:     "enabling the ssh server unprivileged is refused",
+			stored:   &profilemanager.Config{ServerSSHAllowed: boolPtr(false)},
+			change:   privilegedConfigChange{serverSSHAllowed: boolPtr(true)},
+			wantDeny: true,
+		},
+		{
+			name:       "enabling the ssh server as root is allowed",
+			stored:     &profilemanager.Config{ServerSSHAllowed: boolPtr(false)},
+			change:     privilegedConfigChange{serverSSHAllowed: boolPtr(true)},
+			privileged: true,
+		},
+		{
+			name:   "restating an already enabled ssh server is not a change",
+			stored: &profilemanager.Config{ServerSSHAllowed: boolPtr(true)},
+			change: privilegedConfigChange{serverSSHAllowed: boolPtr(true)},
+		},
+		{
+			name:   "turning the ssh server off is not guarded",
+			stored: &profilemanager.Config{ServerSSHAllowed: boolPtr(true)},
+			change: privilegedConfigChange{serverSSHAllowed: boolPtr(false)},
+		},
+		{
+			name:     "a profile with no config yet counts as off, so enabling is refused",
+			stored:   nil,
+			change:   privilegedConfigChange{serverSSHAllowed: boolPtr(true)},
+			wantDeny: true,
+		},
+		{
+			name:     "enabling ssh root login unprivileged is refused",
+			stored:   &profilemanager.Config{EnableSSHRoot: boolPtr(false)},
+			change:   privilegedConfigChange{enableSSHRoot: boolPtr(true)},
+			wantDeny: true,
+		},
+		{
+			name:   "restating ssh root login is not a change",
+			stored: &profilemanager.Config{EnableSSHRoot: boolPtr(true)},
+			change: privilegedConfigChange{enableSSHRoot: boolPtr(true)},
+		},
+		{
+			name:   "turning ssh root login off is not guarded",
+			stored: &profilemanager.Config{EnableSSHRoot: boolPtr(true)},
+			change: privilegedConfigChange{enableSSHRoot: boolPtr(false)},
+		},
+		{
+			name:     "disabling ssh authentication unprivileged is refused",
+			stored:   &profilemanager.Config{DisableSSHAuth: boolPtr(false)},
+			change:   privilegedConfigChange{disableSSHAuth: boolPtr(true)},
+			wantDeny: true,
+		},
+		{
+			name:   "re-enabling ssh authentication is not guarded",
+			stored: &profilemanager.Config{DisableSSHAuth: boolPtr(true)},
+			change: privilegedConfigChange{disableSSHAuth: boolPtr(false)},
+		},
+		{
+			name:   "a request that touches none of the guarded fields is allowed",
+			stored: &profilemanager.Config{ServerSSHAllowed: boolPtr(false)},
+			change: privilegedConfigChange{},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			ctx := userCtx()
+			if tt.privileged {
+				ctx = rootCtx()
+			}
+			err := requirePrivilegeForConfigChange(ctx, tt.stored, tt.change)
+			if tt.wantDeny {
+				assertDenied(t, err)
+				return
+			}
+			assertAllowed(t, err)
+		})
+	}
+}
+
+func TestRequirePrivilegeForConfigChange_ManagementURL(t *testing.T) {
+	sshOn := func(raw string) *profilemanager.Config {
+		return &profilemanager.Config{ServerSSHAllowed: boolPtr(true), ManagementURL: mustURL(t, raw)}
+	}
+	sshOff := func(raw string) *profilemanager.Config {
+		return &profilemanager.Config{ServerSSHAllowed: boolPtr(false), ManagementURL: mustURL(t, raw)}
+	}
+
+	tests := []struct {
+		name       string
+		stored     *profilemanager.Config
+		requested  string
+		privileged bool
+		wantDeny   bool
+	}{
+		{
+			name:      "moving the binding while ssh is enabled is refused",
+			stored:    sshOn("https://api.netbird.io:443"),
+			requested: "https://attacker.example.com:443",
+			wantDeny:  true,
+		},
+		{
+			name:       "moving the binding as root is allowed",
+			stored:     sshOn("https://api.netbird.io:443"),
+			requested:  "https://selfhosted.example.com:443",
+			privileged: true,
+		},
+		{
+			name:      "the same url restated is not a change",
+			stored:    sshOn("https://api.netbird.io:443"),
+			requested: "https://api.netbird.io:443",
+		},
+		{
+			name:      "an equivalent spelling of the same url is not a change",
+			stored:    sshOn("https://api.netbird.io:443"),
+			requested: "https://api.netbird.io",
+		},
+		{
+			name:      "an equivalent spelling with an explicit http port is not a change",
+			stored:    sshOn("http://mgmt.internal:80"),
+			requested: "http://mgmt.internal",
+		},
+		{
+			name:      "a different port on the same host is a change",
+			stored:    sshOn("https://api.netbird.io:443"),
+			requested: "https://api.netbird.io:8443",
+			wantDeny:  true,
+		},
+		{
+			name:      "a different scheme on the same host is a change",
+			stored:    sshOn("https://mgmt.internal:443"),
+			requested: "http://mgmt.internal:443",
+			wantDeny:  true,
+		},
+		{
+			name:      "with ssh disabled the binding is not guarded at all",
+			stored:    sshOff("https://api.netbird.io:443"),
+			requested: "https://attacker.example.com:443",
+		},
+		{
+			name:      "an unparseable url fails closed",
+			stored:    sshOn("https://api.netbird.io:443"),
+			requested: "ht tp://%zz",
+			wantDeny:  true,
+		},
+		{
+			name:      "an empty url leaves the binding alone",
+			stored:    sshOn("https://api.netbird.io:443"),
+			requested: "",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			ctx := userCtx()
+			if tt.privileged {
+				ctx = rootCtx()
+			}
+			err := requirePrivilegeForConfigChange(ctx, tt.stored, privilegedConfigChange{managementURL: tt.requested})
+			if tt.wantDeny {
+				assertDenied(t, err)
+				return
+			}
+			assertAllowed(t, err)
+		})
+	}
+}
+
+// A caller the daemon cannot identify must be refused, not trusted: that is the
+// state on a TCP daemon socket, where no peer credentials exist.
+func TestRequirePrivilegeForConfigChange_UnidentifiedCallerIsRefused(t *testing.T) {
+	err := requirePrivilegeForConfigChange(noIdentityCtx(),
+		&profilemanager.Config{ServerSSHAllowed: boolPtr(false)},
+		privilegedConfigChange{serverSSHAllowed: boolPtr(true)})
+	assertDenied(t, err)
+
+	// The guidance must point at the socket rather than at sudo, since elevating
+	// would not help.
+	st := gstatus.Convert(err)
+	if !strings.Contains(st.Message(), "service install") {
+		t.Errorf("message %q does not tell the operator how to fix the socket", st.Message())
+	}
+}
+
+func TestRequirePrivilegeForDeregistration(t *testing.T) {
+	tests := []struct {
+		name       string
+		cfg        *profilemanager.Config
+		privileged bool
+		wantDeny   bool
+	}{
+		{
+			name:     "deregistering while ssh is enabled is refused",
+			cfg:      &profilemanager.Config{ServerSSHAllowed: boolPtr(true)},
+			wantDeny: true,
+		},
+		{
+			name:       "deregistering while ssh is enabled is allowed for root",
+			cfg:        &profilemanager.Config{ServerSSHAllowed: boolPtr(true)},
+			privileged: true,
+		},
+		{
+			name: "deregistering with ssh disabled is not guarded",
+			cfg:  &profilemanager.Config{ServerSSHAllowed: boolPtr(false)},
+		},
+		{
+			name: "deregistering a profile with no config is not guarded",
+			cfg:  nil,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			ctx := userCtx()
+			if tt.privileged {
+				ctx = rootCtx()
+			}
+			err := requirePrivilegeForDeregistration(ctx, tt.cfg)
+			if tt.wantDeny {
+				assertDenied(t, err)
+				return
+			}
+			assertAllowed(t, err)
+		})
+	}
+}
+
+// privilegedTestCtx is the context a handler-level test should use when it is
+// standing in for a root/administrator caller. Tests that drive the handlers
+// directly have no transport credentials, and the privileged-change gate refuses
+// a caller it cannot identify.
+func privilegedTestCtx() context.Context { return rootCtx() }

client/ssh/client/client.go

@@ -9,7 +9,6 @@ import (
 	"path/filepath"
 	"runtime"
 	"strconv"
-	"strings"
 	"time"
 
 	log "github.com/sirupsen/logrus"
@@ -17,7 +16,6 @@ import (
 	"golang.org/x/crypto/ssh/knownhosts"
 	"golang.org/x/term"
 	"google.golang.org/grpc"
-	"google.golang.org/grpc/credentials/insecure"
 
 	"github.com/netbirdio/netbird/client/internal/daemonaddr"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
@@ -32,7 +30,7 @@ const (
 	// DefaultDaemonAddr is the default address for the NetBird daemon
 	DefaultDaemonAddr = "unix:///var/run/netbird.sock"
 	// DefaultDaemonAddrWindows is the default address for the NetBird daemon on Windows
-	DefaultDaemonAddrWindows = "tcp://127.0.0.1:41731"
+	DefaultDaemonAddrWindows = daemonaddr.WindowsPipeAddr
 )
 
 // Client wraps crypto/ssh Client for simplified SSH operations
@@ -268,7 +266,7 @@ func getDefaultDaemonAddr() string {
 		return addr
 	}
 	if runtime.GOOS == "windows" {
-		return DefaultDaemonAddrWindows
+		return daemonaddr.ResolveDaemonAddr(DefaultDaemonAddrWindows)
 	}
 	return daemonaddr.ResolveUnixDaemonAddr(DefaultDaemonAddr)
 }
@@ -410,12 +408,9 @@ func verifyHostKeyViaDaemon(hostname string, remote net.Addr, key ssh.PublicKey,
 }
 
 func connectToDaemon(daemonAddr string) (*grpc.ClientConn, error) {
-	addr := strings.TrimPrefix(daemonAddr, "tcp://")
+	target, opts := daemonaddr.DialTarget(daemonAddr)
 
-	conn, err := grpc.NewClient(
-		addr,
-		grpc.WithTransportCredentials(insecure.NewCredentials()),
-	)
+	conn, err := grpc.NewClient(target, opts...)
 	if err != nil {
 		log.Debugf("failed to create gRPC client for NetBird daemon at %s: %v", daemonAddr, err)
 		return nil, fmt.Errorf("failed to connect to NetBird daemon: %w", err)

client/ssh/proxy/proxy.go

@@ -9,7 +9,6 @@ import (
 	"net"
 	"os"
 	"strconv"
-	"strings"
 	"sync"
 	"time"
 
@@ -17,8 +16,8 @@ import (
 	log "github.com/sirupsen/logrus"
 	cryptossh "golang.org/x/crypto/ssh"
 	"google.golang.org/grpc"
-	"google.golang.org/grpc/credentials/insecure"
 
+	"github.com/netbirdio/netbird/client/internal/daemonaddr"
 	"github.com/netbirdio/netbird/client/internal/profilemanager"
 	"github.com/netbirdio/netbird/client/proto"
 	nbssh "github.com/netbirdio/netbird/client/ssh"
@@ -55,8 +54,8 @@ type SSHProxy struct {
 }
 
 func New(daemonAddr, targetHost string, targetPort int, stderr io.Writer, browserOpener func(string) error) (*SSHProxy, error) {
-	grpcAddr := strings.TrimPrefix(daemonAddr, "tcp://")
-	grpcConn, err := grpc.NewClient(grpcAddr, grpc.WithTransportCredentials(insecure.NewCredentials()))
+	target, opts := daemonaddr.DialTarget(daemonAddr)
+	grpcConn, err := grpc.NewClient(target, opts...)
 	if err != nil {
 		return nil, fmt.Errorf("connect to daemon: %w", err)
 	}

client/ui/frontend/src/components/CopyToClipboard.tsx

@@ -18,6 +18,9 @@ type CopyToClipboardProps = {
     className?: string;
     iconClassName?: string;
     alwaysShowIcon?: boolean;
+    // wrap lets long content (a shell command, a path) break across lines
+    // instead of being truncated to one line.
+    wrap?: boolean;
     variant?: CopyToClipboardVariant;
     "aria-label"?: string;
     tabIndex?: number;
@@ -32,6 +35,7 @@ export const CopyToClipboard = ({
     className,
     iconClassName,
     alwaysShowIcon = false,
+    wrap = false,
     variant = "default",
     "aria-label": ariaLabel,
     tabIndex = 0,
@@ -83,7 +87,8 @@ export const CopyToClipboard = ({
         >
             <span
                 className={cn(
-                    "relative min-w-0 truncate",
+                    "relative min-w-0",
+                    wrap ? "whitespace-pre-wrap break-all" : "truncate",
                     "[&_*]:transition-colors",
                     VARIANT_HOVER[variant],
                 )}

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

@@ -14,7 +14,7 @@ import type { Config } from "@bindings/services/models.js";
 import i18next from "@/lib/i18n";
 import { useProfile } from "@/contexts/ProfileContext.tsx";
 import { SettingsSkeleton } from "@/modules/settings/SettingsSkeleton.tsx";
-import { errorDialog, formatErrorMessage as errorMessage } from "@/lib/errors.ts";
+import { errorCommand, errorDialog, formatErrorMessage as errorMessage } from "@/lib/errors.ts";
 
 const SAVE_DEBOUNCE_MS = 400;
 
@@ -68,6 +68,21 @@ const useSettingsState = () => {
         loadedRef.current = loaded;
     }, [loaded]);
 
+    // reload re-reads the daemon's config, which is authoritative. Used on
+    // mount, on the daemon's config_changed event, and to undo an optimistic
+    // update the daemon then rejected.
+    const reload = useCallback(
+        async (profileName: string) => {
+            try {
+                const data = await SettingsSvc.GetConfig({ profileName, username });
+                setLoaded({ profileName, data });
+            } catch (e) {
+                console.warn("[SettingsContext] reload after rejected save failed", e);
+            }
+        },
+        [username],
+    );
+
     useEffect(() => {
         if (!profileLoaded || !activeProfileId) return;
         let cancelled = false;
@@ -133,13 +148,20 @@ const useSettingsState = () => {
                     username,
                 });
             } catch (e) {
+                // The optimistic update is wrong now: the daemon refused it
+                // (a change that needs elevated privileges, an MDM-managed
+                // field, ...). Snap the controls back to what it actually
+                // holds before reporting, so the UI never shows a value the
+                // daemon does not have.
+                await reload(profileName);
                 await errorDialog({
                     Title: i18next.t("settings.error.saveTitle"),
                     Message: errorMessage(e),
+                    Command: errorCommand(e),
                 });
             }
         },
-        [username],
+        [username, reload],
     );
 
     const setField = useCallback(

client/ui/frontend/src/hooks/usePrivilege.ts

@@ -0,0 +1,32 @@
+import { useEffect, useState } from "react";
+import { Settings as SettingsSvc } from "@bindings/services";
+import { Privilege } from "@bindings/services/models.js";
+
+// usePrivilege reports whether this UI process may perform the changes the daemon
+// restricts to root/administrator. It is answered in-process from our own token
+// with the daemon's own rule, so there is no round-trip and it works while the
+// daemon is down.
+//
+// null means "not known yet", which includes the read having failed. Callers must
+// treat that as "do not restrict": the daemon enforces this regardless, so the
+// only thing a wrong guess here costs is a control that looks unavailable when it
+// is not, or a save that fails with the daemon's own guidance.
+export const usePrivilege = (): Privilege | null => {
+    const [privilege, setPrivilege] = useState<Privilege | null>(null);
+
+    useEffect(() => {
+        let cancelled = false;
+        SettingsSvc.Privilege()
+            .then((p) => {
+                if (!cancelled) setPrivilege(p);
+            })
+            .catch((e: unknown) => {
+                console.warn("[usePrivilege] read failed, not restricting controls", e);
+            });
+        return () => {
+            cancelled = true;
+        };
+    }, []);
+
+    return privilege;
+};

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

@@ -1,6 +1,6 @@
 import { WindowManager } from "@bindings/services";
 
-type ClassifiedError = { short: string; long: string };
+type ClassifiedError = { short: string; long: string; command: string };
 
 const asObject = (v: unknown): Record<string, unknown> | null =>
     v && typeof v === "object" ? (v as Record<string, unknown>) : null;
@@ -22,20 +22,24 @@ const toWailsEnvelope = (e: unknown): Record<string, unknown> | null => {
     return asObject(obj.cause) ?? parseJsonObject(obj.message);
 };
 
-// Read { short, long } from wherever the classified error sits in the envelope
+// Read { short, long, command } from wherever the classified error sits in the envelope
 const toClassifiedError = (v: unknown): ClassifiedError | null => {
     const o = asObject(v);
     if (!o) return null;
     const short = typeof o.short === "string" ? o.short : "";
     const long = typeof o.long === "string" ? o.long : "";
-    return short || long ? { short, long } : null;
+    const command = typeof o.command === "string" ? o.command : "";
+    return short || long ? { short, long, command } : null;
+};
+
+const classify = (e: unknown): ClassifiedError | null => {
+    const envelope = toWailsEnvelope(e);
+    return toClassifiedError(envelope?.cause) ?? toClassifiedError(envelope);
 };
 
 export const formatErrorMessage = (e: unknown): string => {
-    const envelope = toWailsEnvelope(e);
-
     // Prefer the structured { short, long } the daemon classifier produced.
-    const classified = toClassifiedError(envelope?.cause) ?? toClassifiedError(envelope);
+    const classified = classify(e);
     if (classified) {
         const { short, long } = classified;
         if (short && long && long !== short) return `${short} Details: ${long}`;
@@ -44,17 +48,26 @@ export const formatErrorMessage = (e: unknown): string => {
     }
 
     // Unclassified (a service returned the raw daemon error)
+    const envelope = toWailsEnvelope(e);
     const message = envelope?.message;
     if (typeof message === "string" && message) return message;
     if (e instanceof Error) return e.message;
     return String(e);
 };
 
+// errorCommand returns a command the user can run to complete an operation the
+// daemon refused, when the error carries one (a change that needs elevated
+// privileges). Empty for every other error.
+export const errorCommand = (e: unknown): string => classify(e)?.command ?? "";
+
 export type ErrorDialogOptions = {
     Title: string;
     Message: string;
+    // Command is shown for copying below the message. Defaults to the one the
+    // error carries, so callers only pass it to override.
+    Command?: string;
 };
 
 export function errorDialog(options: ErrorDialogOptions): Promise<void> {
-    return WindowManager.OpenError(options.Title, options.Message);
+    return WindowManager.OpenError(options.Title, options.Message, options.Command ?? "");
 }

client/ui/frontend/src/modules/error/ErrorDialog.tsx

@@ -3,6 +3,7 @@ import { useTranslation } from "react-i18next";
 import { useSearchParams } from "react-router-dom";
 import { AlertCircleIcon } from "lucide-react";
 import { Button } from "@/components/buttons/Button";
+import { CopyToClipboard } from "@/components/CopyToClipboard";
 import { ConfirmDialog } from "@/components/dialog/ConfirmDialog";
 import { DialogActions } from "@/components/dialog/DialogActions";
 import { DialogDescription } from "@/components/dialog/DialogDescription";
@@ -12,14 +13,22 @@ import { WindowManager } from "@bindings/services";
 import { useAutoSizeWindow } from "@/hooks/useAutoSizeWindow";
 
 const WINDOW_WIDTH = 380;
+// A command needs the room to wrap at a sensible number of characters instead of
+// breaking every few words.
+const WINDOW_WIDTH_WITH_COMMAND = 460;
 
 export default function ErrorDialog() {
     const { t } = useTranslation();
-    const contentRef = useAutoSizeWindow<HTMLDivElement>(WINDOW_WIDTH);
     const [params] = useSearchParams();
 
     const title = params.get("title") || t("window.title.error");
     const message = params.get("message") || "";
+    // Set when the daemon refused an operation that needs elevated privileges:
+    // the command that performs it, offered for copying.
+    const command = params.get("command") || "";
+    const contentRef = useAutoSizeWindow<HTMLDivElement>(
+        command ? WINDOW_WIDTH_WITH_COMMAND : WINDOW_WIDTH,
+    );
 
     const close = useCallback(() => {
         WindowManager.CloseError().catch(console.error);
@@ -37,15 +46,37 @@ export default function ErrorDialog() {
         <ConfirmDialog ref={contentRef} aria-labelledby={"nb-error-dialog-title"}>
             <SquareIcon icon={AlertCircleIcon} variant={"danger"} />
 
-            <div className={"flex flex-col items-center gap-1"}>
+            <div className={"flex w-full flex-col items-center gap-1"}>
                 <DialogHeading id={"nb-error-dialog-title"} className={"text-balance"}>
                     {title}
                 </DialogHeading>
                 {message && (
                     <DialogDescription className={"text-balance"}>
-                        <span className={"whitespace-pre-wrap break-words"}>{message}</span>
+                        {/* select-text: the message often names a path, a flag or an
+                            address the user needs to act on. */}
+                        <span className={"select-text whitespace-pre-wrap break-words"}>
+                            {message}
+                        </span>
                     </DialogDescription>
                 )}
+                {command && (
+                    <CopyToClipboard
+                        message={command}
+                        alwaysShowIcon
+                        wrap
+                        variant={"bright"}
+                        className={
+                            "mt-2 w-full items-start gap-2 rounded-md bg-nb-gray-930 px-3 py-2 text-left"
+                        }
+                        aria-label={t("common.copy")}
+                    >
+                        <code
+                            className={"select-text break-all font-mono text-xs text-nb-gray-200"}
+                        >
+                            {command}
+                        </code>
+                    </CopyToClipboard>
+                )}
             </div>
 
             <DialogActions>

client/ui/frontend/src/modules/settings/SettingsSSH.tsx

@@ -1,4 +1,5 @@
 import { useTranslation } from "react-i18next";
+import { CopyToClipboard } from "@/components/CopyToClipboard";
 import FancyToggleSwitch from "@/components/switches/FancyToggleSwitch";
 import { HelpText } from "@/components/typography/HelpText";
 import { Input } from "@/components/inputs/Input";
@@ -6,12 +7,50 @@ import { Label } from "@/components/typography/Label";
 import { cn } from "@/lib/cn";
 import { SectionGroup } from "@/modules/settings/SettingsSection.tsx";
 import { useSettings } from "@/contexts/SettingsContext.tsx";
-import { type ChangeEvent, useEffect, useId, useState } from "react";
+import { usePrivilege } from "@/hooks/usePrivilege.ts";
+import { Privilege } from "@bindings/services/models.js";
+import { type ChangeEvent, type ReactNode, useEffect, useId, useState } from "react";
 
 export function SettingsSSH() {
     const { t } = useTranslation();
     const { config, setField } = useSettings();
+    const privilege = usePrivilege();
     const isSSHServerEnabled = config.serverSshAllowed;
+
+    // The daemon restricts only the direction that hands out shells from a process
+    // running as root. So for an unprivileged user a guarded control is either
+    // unavailable (it is off and only they could turn it on) or a one-way switch
+    // (it is on, they may turn it off, but not back on) — say which, either way.
+    //
+    // A null privilege means we could not determine it: leave the control alone
+    // rather than greying it out with nothing to explain why. The daemon enforces
+    // this regardless, and a rejected save reports its own guidance.
+    const guarded = (
+        guardedDirectionActive: boolean,
+        command: (p: Privilege) => string,
+        // inverted marks a control whose guarded direction is switching it off, so
+        // the one-way warning has to read the other way round.
+        inverted = false,
+    ) => {
+        if (!privilege || privilege.privileged) {
+            return { disabled: false, hint: undefined };
+        }
+        const hint = (
+            <PrivilegeHint
+                actor={privilege.actor}
+                command={command(privilege)}
+                oneWay={guardedDirectionActive}
+                inverted={inverted}
+            />
+        );
+        return { disabled: !guardedDirectionActive, hint };
+    };
+
+    const sshServer = guarded(config.serverSshAllowed, (p) => p.allowSshServer);
+    const sshRoot = guarded(config.enableSshRoot, (p) => p.enableSshRoot);
+    // Inverted control: the guarded direction is switching authentication off, so
+    // it is the already-disabled state that is the one-way one.
+    const sshAuth = guarded(config.disableSshAuth, (p) => p.disableSshAuth, true);
     const jwtTtlId = useId();
     const [jwtTtlInput, setJwtTtlInput] = useState(String(config.sshJwtCacheTtl));
 
@@ -46,9 +85,11 @@ export function SettingsSSH() {
                 <FancyToggleSwitch
                     value={config.serverSshAllowed}
                     onChange={(v) => setField("serverSshAllowed", v)}
+                    disabled={sshServer.disabled}
                     label={t("settings.ssh.server.label")}
                     helpText={t("settings.ssh.server.help")}
                 />
+                {sshServer.hint}
             </SectionGroup>
 
             <SectionGroup
@@ -58,9 +99,11 @@ export function SettingsSSH() {
                 <FancyToggleSwitch
                     value={config.enableSshRoot}
                     onChange={(v) => setField("enableSshRoot", v)}
+                    disabled={sshRoot.disabled}
                     label={t("settings.ssh.root.label")}
                     helpText={t("settings.ssh.root.help")}
                 />
+                {sshRoot.hint}
                 <FancyToggleSwitch
                     value={config.enableSshSftp}
                     onChange={(v) => setField("enableSshSftp", v)}
@@ -88,9 +131,11 @@ export function SettingsSSH() {
                 <FancyToggleSwitch
                     value={!config.disableSshAuth}
                     onChange={(v) => setField("disableSshAuth", !v)}
+                    disabled={sshAuth.disabled}
                     label={t("settings.ssh.jwt.label")}
                     helpText={t("settings.ssh.jwt.help")}
                 />
+                {sshAuth.hint}
                 <div
                     className={cn(
                         "flex items-center justify-between gap-6",
@@ -117,3 +162,42 @@ export function SettingsSSH() {
         </>
     );
 }
+
+// PrivilegeHint explains what an unprivileged user can and cannot do with a
+// guarded control, and offers the command that does it with the privileges the
+// daemon requires. oneWay covers the control being in the guarded state already:
+// switching it back is the part that needs privileges.
+function PrivilegeHint({
+    actor,
+    command,
+    oneWay,
+    inverted,
+}: {
+    actor: string;
+    command: string;
+    oneWay: boolean;
+    inverted: boolean;
+}): ReactNode {
+    const { t } = useTranslation();
+    if (!command) return null;
+    return (
+        <div
+            className={
+                "-mt-2 flex flex-col gap-1 rounded-md bg-nb-gray-930 px-3 py-2 text-xs text-nb-gray-300"
+            }
+        >
+            <span>
+                {!oneWay
+                    ? t("settings.ssh.privilege.hint", { actor })
+                    : inverted
+                      ? t("settings.ssh.privilege.oneWayInverted", { actor })
+                      : t("settings.ssh.privilege.oneWay", { actor })}
+            </span>
+            <CopyToClipboard message={command} alwaysShowIcon wrap variant={"bright"}>
+                <code className={"select-text break-all font-mono text-xs text-nb-gray-200"}>
+                    {command}
+                </code>
+            </CopyToClipboard>
+        </div>
+    );
+}

client/ui/grpc.go

@@ -5,14 +5,13 @@ package main
 import (
 	"fmt"
 	"runtime"
-	"strings"
 	"sync"
 	"time"
 
 	"google.golang.org/grpc"
 	"google.golang.org/grpc/backoff"
-	"google.golang.org/grpc/credentials/insecure"
 
+	"github.com/netbirdio/netbird/client/internal/daemonaddr"
 	"github.com/netbirdio/netbird/client/proto"
 	"github.com/netbirdio/netbird/client/ui/desktop"
 )
@@ -36,9 +35,10 @@ func (c *Conn) Client() (proto.DaemonServiceClient, error) {
 		return c.client, nil
 	}
 
-	cc, err := grpc.NewClient(
-		strings.TrimPrefix(c.addr, "tcp://"),
-		grpc.WithTransportCredentials(insecure.NewCredentials()),
+	// Lazy on purpose: grpc.NewClient does not connect here, so a daemon that
+	// is down surfaces as a per-RPC Unavailable instead of blocking the UI.
+	target, opts := daemonaddr.DialTarget(daemonaddr.ResolveDaemonAddr(c.addr))
+	opts = append(opts,
 		grpc.WithUserAgent(desktop.GetUIUserAgent()),
 		// Cap reconnect backoff at 5s; gRPC's default 120s MaxDelay would
 		// leave the UI waiting 30-60s to notice a freshly-started daemon.
@@ -51,6 +51,8 @@ func (c *Conn) Client() (proto.DaemonServiceClient, error) {
 			},
 		}),
 	)
+
+	cc, err := grpc.NewClient(target, opts...)
 	if err != nil {
 		return nil, fmt.Errorf("dial daemon: %w", err)
 	}
@@ -58,10 +60,12 @@ func (c *Conn) Client() (proto.DaemonServiceClient, error) {
 	return c.client, nil
 }
 
-// DaemonAddr returns the default daemon gRPC address: a Unix socket on Linux/macOS, TCP loopback on Windows.
+// DaemonAddr returns the default daemon gRPC address: a Unix socket on
+// Linux/macOS, a named pipe on Windows. The pipe carries the caller's token,
+// which loopback TCP does not, so the daemon can tell who is calling.
 func DaemonAddr() string {
 	if runtime.GOOS == "windows" {
-		return "tcp://127.0.0.1:41731"
+		return daemonaddr.WindowsPipeAddr
 	}
 	return "unix:///var/run/netbird.sock"
 }

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

@@ -1774,5 +1774,17 @@
     "error.unknown": {
         "message": "Operation failed.",
         "description": "Generic fallback error message used when no specific error applies."
+    },
+    "settings.ssh.privilege.hint": {
+        "message": "Requires {actor}. Run this instead:",
+        "description": "Help text under an SSH setting the user cannot change: it needs elevated privileges. {actor} is 'root' on Linux/macOS or 'administrator privileges' on Windows. Followed by a copyable command."
+    },
+    "settings.ssh.privilege.oneWay": {
+        "message": "You can switch this off, but switching it back on needs {actor}:",
+        "description": "Warning under an SSH setting an unprivileged user may disable but not re-enable. {actor} is 'root' on Linux/macOS or 'administrator privileges' on Windows. Followed by a copyable command."
+    },
+    "settings.ssh.privilege.oneWayInverted": {
+        "message": "You can switch this on, but switching it back off needs {actor}:",
+        "description": "Warning under the SSH authentication setting, which an unprivileged user may re-enable but not disable again. {actor} is 'root' on Linux/macOS or 'administrator privileges' on Windows. Followed by a copyable command."
     }
 }

client/ui/main.go

@@ -96,7 +96,6 @@ func main() {
 		}
 	})
 
-	settings := services.NewSettings(conn)
 	profiles := services.NewProfiles(conn)
 	// updater.Holder owns the typed update State; DaemonFeed feeds it and the
 	// Update service is a thin Wails-bound facade over it plus the install RPCs.
@@ -117,6 +116,7 @@ func main() {
 	bundle, prefStore, localizer := buildI18n(app)
 
 	// After bundle + prefStore: both are used to localise daemon errors.
+	settings := services.NewSettings(conn, bundle, prefStore, daemonAddr)
 	connection := services.NewConnection(conn, bundle, prefStore)
 	profileSwitcher := services.NewProfileSwitcher(profiles, connection, daemonFeed)
 	// authsession.Session owns the full extend + dismiss surface the tray

client/ui/services/errors.go

@@ -6,13 +6,30 @@ import (
 	"encoding/json"
 	"strings"
 
+	"google.golang.org/genproto/googleapis/rpc/errdetails"
 	gcodes "google.golang.org/grpc/codes"
 	gstatus "google.golang.org/grpc/status"
 
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
 	"github.com/netbirdio/netbird/client/ui/i18n"
 	"github.com/netbirdio/netbird/client/ui/preferences"
 )
 
+// privilegeErrorInfo returns the daemon's privilege-refusal detail, if the error
+// carries one.
+func privilegeErrorInfo(err error) (*errdetails.ErrorInfo, bool) {
+	for _, detail := range gstatus.Convert(err).Details() {
+		info, ok := detail.(*errdetails.ErrorInfo)
+		if !ok {
+			continue
+		}
+		if info.GetReason() == ipcauth.ErrorReasonPrivilegeRequired && info.GetDomain() == ipcauth.ErrorDomain {
+			return info, true
+		}
+	}
+	return nil, false
+}
+
 // ErrorTranslator localises daemon errors; runtime impl is *i18n.Bundle.
 type ErrorTranslator interface {
 	Translate(lang i18n.LanguageCode, key string, args ...string) string
@@ -30,6 +47,10 @@ type ClientError struct {
 	Code  string `json:"code"`
 	Short string `json:"short"`
 	Long  string `json:"long"`
+	// Command is a command the user can run to complete the operation
+	// themselves, set when the daemon refused it for want of privileges. The
+	// frontend offers it for copying.
+	Command string `json:"command,omitempty"`
 }
 
 // Error returns the short message for plain Go callers.
@@ -72,6 +93,24 @@ func (c errorClassifier) classify(err error) *ClientError {
 		msg = st.Message()
 		grpcCode = st.Code()
 	}
+
+	// A refusal for want of privileges carries its own summary and the command
+	// that performs the operation, both written for the user. Surface them
+	// verbatim: no substring guessing, and no localisation of a message the
+	// daemon composed.
+	if info, ok := privilegeErrorInfo(err); ok {
+		summary := info.GetMetadata()[ipcauth.ErrorMetaSummary]
+		if summary == "" {
+			summary = msg
+		}
+		return &ClientError{
+			Code:    "privilege_required",
+			Short:   summary,
+			Long:    summary,
+			Command: info.GetMetadata()[ipcauth.ErrorMetaCommand],
+		}
+	}
+
 	lower := strings.ToLower(msg)
 
 	code := "unknown"

client/ui/services/settings.go

@@ -7,6 +7,10 @@ import (
 	"fmt"
 	"reflect"
 
+	log "github.com/sirupsen/logrus"
+
+	"github.com/netbirdio/netbird/client/internal/daemonaddr"
+	"github.com/netbirdio/netbird/client/internal/ipcauth"
 	"github.com/netbirdio/netbird/client/proto"
 )
 
@@ -39,6 +43,19 @@ type Restrictions struct {
 	Features Features  `json:"features"`
 }
 
+// Privilege tells the frontend whether this process may perform the changes the
+// daemon restricts to root/administrator, and carries the command for each so a
+// disabled control can show the way to do it.
+type Privilege struct {
+	Privileged bool `json:"privileged"`
+	// Actor names what the operation requires ("root", "administrator privileges").
+	Actor string `json:"actor"`
+	// Commands equivalent to the settings the daemon guards, ready to copy.
+	AllowSSHServer string `json:"allowSshServer"`
+	EnableSSHRoot  string `json:"enableSshRoot"`
+	DisableSSHAuth string `json:"disableSshAuth"`
+}
+
 type ConfigParams struct {
 	ProfileName string `json:"profileName"`
 	Username    string `json:"username"`
@@ -106,11 +123,19 @@ type SetConfigParams struct {
 }
 
 type Settings struct {
-	conn DaemonConn
+	conn       DaemonConn
+	classifier errorClassifier
+	// daemonAddr is where the daemon listens, used to tell whether it runs as
+	// this user and would therefore authorize us: see Privilege.
+	daemonAddr string
 }
 
-func NewSettings(conn DaemonConn) *Settings {
-	return &Settings{conn: conn}
+func NewSettings(conn DaemonConn, translator ErrorTranslator, prefs LanguagePreference, daemonAddr string) *Settings {
+	return &Settings{
+		conn:       conn,
+		classifier: errorClassifier{translator: translator, prefs: prefs},
+		daemonAddr: daemonAddr,
+	}
 }
 
 func (s *Settings) GetConfig(ctx context.Context, p ConfigParams) (Config, error) {
@@ -189,8 +214,47 @@ func (s *Settings) SetConfig(ctx context.Context, p SetConfigParams) error {
 		DisableSSHAuth:                p.DisableSSHAuth,
 		SshJWTCacheTTL:                p.SSHJWTCacheTTL,
 	}
-	_, err = cli.SetConfig(ctx, req)
-	return err
+	if _, err := cli.SetConfig(ctx, req); err != nil {
+		// Classified so the frontend gets the daemon's guidance instead of the
+		// gRPC envelope, which is what a refused privileged change looks like.
+		return s.classifier.classify(err)
+	}
+	return nil
+}
+
+// Privilege reports whether this UI process could carry out the changes the
+// daemon restricts to root/administrator, and the command that performs the one
+// users hit in the SSH settings. It applies the daemon's own rule to what it can
+// see locally, so the frontend can present those controls as unavailable up front
+// instead of letting a save fail. No daemon round-trip, so it also works while the
+// daemon is down.
+//
+// Being root or an elevated administrator is one way. The other is running as the
+// daemon's own user while the daemon is unprivileged, which the daemon accepts
+// because such a caller can already rewrite the config it reads; that is the
+// rootless-container and Windows netstack-mode case, and it is read from the
+// ownership of the socket or pipe the daemon created.
+func (s *Settings) Privilege() Privilege {
+	id, err := ipcauth.CurrentProcessIdentity()
+	if err != nil {
+		// Fail closed: report unprivileged, which only ever disables controls.
+		log.Warnf("cannot read this process's identity, treating it as unprivileged: %v", err)
+		return newPrivilege(false)
+	}
+	if id.IsPrivileged() {
+		return newPrivilege(true)
+	}
+	return newPrivilege(daemonaddr.DaemonRunsAsSelf(s.daemonAddr))
+}
+
+func newPrivilege(privileged bool) Privilege {
+	return Privilege{
+		Privileged:     privileged,
+		Actor:          ipcauth.PrivilegedActor(),
+		AllowSSHServer: ipcauth.UpCommand("--allow-server-ssh"),
+		EnableSSHRoot:  ipcauth.UpCommand("--enable-ssh-root"),
+		DisableSSHAuth: ipcauth.UpCommand("--disable-ssh-auth"),
+	}
 }
 
 func (s *Settings) GetRestrictions(ctx context.Context) (Restrictions, error) {

client/ui/services/windowmanager.go

@@ -393,15 +393,17 @@ func (s *WindowManager) CloseWelcome() {
 	}
 }
 
-// OpenError shows the custom error dialog; title/message are pre-localised and ride in the
-// start URL. A second error replaces the open one via SetURL. Singleton, destroyed on close.
-func (s *WindowManager) OpenError(title, message string) {
+// OpenError shows the custom error dialog; title/message/command are pre-localised
+// and ride in the start URL. command is optional and, when set, is offered for
+// copying so the user can run the operation the daemon refused. A second error
+// replaces the open one via SetURL. Singleton, destroyed on close.
+func (s *WindowManager) OpenError(title, message, command string) {
 	if ShuttingDown() {
 		return
 	}
 	s.mu.Lock()
 	defer s.mu.Unlock()
-	startURL := errorDialogURL(title, message)
+	startURL := errorDialogURL(title, message, command)
 	if s.errorDialog == nil {
 		s.errorDialog = s.app.Window.NewWithOptions(
 			DialogWindowOptions("error", s.title("window.title.error"), startURL, s.linuxIcon),
@@ -601,8 +603,8 @@ func (s *WindowManager) getScreenBasedOnCursorPosition() *application.Screen {
 	return nil
 }
 
-// errorDialogURL builds the error window's start URL with title/message as escaped query params.
-func errorDialogURL(title, message string) string {
+// errorDialogURL builds the error window's start URL with title/message/command as escaped query params.
+func errorDialogURL(title, message, command string) string {
 	q := url.Values{}
 	if title != "" {
 		q.Set("title", title)
@@ -610,6 +612,9 @@ func errorDialogURL(title, message string) string {
 	if message != "" {
 		q.Set("message", message)
 	}
+	if command != "" {
+		q.Set("command", command)
+	}
 	startURL := "/#/dialog/error"
 	if enc := q.Encode(); enc != "" {
 		startURL += "?" + enc

go.mod

@@ -30,6 +30,7 @@ require (
 
 require (
 	github.com/DeRuina/timberjack v1.4.2
+	github.com/Microsoft/go-winio v0.6.2
 	github.com/awnumar/memguard v0.23.0
 	github.com/aws/aws-sdk-go-v2 v1.38.3
 	github.com/aws/aws-sdk-go-v2/aws/protocol/eventstream v1.7.1
@@ -133,6 +134,7 @@ require (
 	golang.org/x/term v0.45.0
 	golang.org/x/time v0.15.0
 	google.golang.org/api v0.276.0
+	google.golang.org/genproto/googleapis/rpc v0.0.0-20260401024825-9d38bb4040a9
 	gopkg.in/yaml.v3 v3.0.1
 	gorm.io/driver/mysql v1.5.7
 	gorm.io/driver/postgres v1.5.7
@@ -156,7 +158,6 @@ require (
 	github.com/Masterminds/goutils v1.1.1 // indirect
 	github.com/Masterminds/semver/v3 v3.4.0 // indirect
 	github.com/Masterminds/sprig/v3 v3.3.0 // indirect
-	github.com/Microsoft/go-winio v0.6.2 // indirect
 	github.com/adrg/xdg v0.5.3 // indirect
 	github.com/anmitsu/go-shlex v0.0.0-20200514113438-38f4b401e2be // indirect
 	github.com/apapsch/go-jsonmerge/v2 v2.0.0 // indirect
@@ -317,7 +318,6 @@ require (
 	golang.org/x/tools v0.47.0 // indirect
 	golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2 // indirect
 	google.golang.org/genproto/googleapis/api v0.0.0-20260319201613-d00831a3d3e7 // indirect
-	google.golang.org/genproto/googleapis/rpc v0.0.0-20260401024825-9d38bb4040a9 // indirect
 	gopkg.in/square/go-jose.v2 v2.6.0 // indirect
 	gopkg.in/tomb.v1 v1.0.0-20141024135613-dd632973f1e7 // indirect
 	gopkg.in/yaml.v2 v2.4.0 // indirect