GHSA-67jx-r9pv-98rj

Summary

Traefik's Knative provider builds router rules by interpolating user-controlled values into backtick-delimited rule expressions without escaping. In live cluster validation, Knative rules[].hosts[] was exploitable for host restriction bypass (for example tenant.example.com) || Host(attacker.com), producing a router that serves attacker-controlled hosts. Knative headers[].exact also allows rule-syntax injection and proves unsafe rule construction. In multi-tenant clusters, this can route unauthorized traffic to victim services and lead to cross-tenant traffic exposure. Severity is High in shared deployments.

Tested on Traefik v3.6.10; the vulnerable pattern appears to have been present since the Knative provider was introduced. Earlier versions with Knative provider support are expected to be affected.

Details

The issue is caused by unsafe rule-string construction using fmt.Sprintf with backtick-delimited literals.

Incriminated code patterns:

• pkg/provider/kubernetes/knative/kubernetes.go

  • fmt.Sprintf("Host(%v)", host)
  • fmt.Sprintf("Header(%s,%s)", key, headers[key].Exact)
  • fmt.Sprintf("PathPrefix(%s)", path)

• pkg/provider/kubernetes/ingress/kubernetes.go

  • fmt.Sprintf("Host(%s)", host)
  • fmt.Sprintf("(Path(%[1]s) || PathPrefix(%[1]s/))", path)

• pkg/provider/kubernetes/ingress-nginx/kubernetes.go (hardening candidate; not the primary confirmed vector in this report)

  • fmt.Sprintf("Header(%s, %s)", c.Header, c.HeaderValue)
  • related host/path/header concatenations with backticks

Because inputs are inserted directly into rule expressions, a malicious value containing a backtick can terminate the literal and inject additional operators/tokens in Traefik's rule language. Example payload:

• x) || Host(attacker.com

When used as a header value in Knative rule construction, the resulting rule contains:

• Header(X-Poc,x) || Host(attacker.com)

This alters rule semantics and enables injection into Traefik's rule language. Depending on the field used (hosts[] vs headers[].exact) this can become a direct routing bypass.

Important scope note:

• Gateway API code path (pkg/provider/kubernetes/gateway/httproute.go) already uses safer %q formatting for header/query rules and is not affected by this exact pattern.
• For standard Kubernetes Ingress, spec.rules.host is validated as DNS-1123 by the API server, which rejects backticks (so this specific host-injection payload is typically blocked).
• For Knative Ingress, rules[].hosts[] and headers[].exact are typed as string in CRD schema with no pattern constraint.
• In this validation environment, rules[].hosts[] was accepted and produced a practical host bypass. headers[].exact was also accepted and produced rule-syntax injection in generated routers.
• Ingress-NGINX patterns are included as follow-up hardening targets and are not claimed as independently exploitable here.
• Exploitability depends on admission/validation policy and who can create these resources.

PoC

1. Local deterministic PoC (no cluster required):

• Run:
  • Save the inline PoC below as poc_build_rule.go
  • Run go run poc_build_rule.go
• Observe output:
  • Legitimate rule: (Host(tenant.example.com)) && (Header(X-API-Key,secret123)) && PathPrefix(/)
  • Malicious rule: (Host(tenant.example.com)) && (Header(X-API-Key,x) || Host(attacker.com)) && PathPrefix(/)
• This proves syntax injection in current string-construction logic.

Inline PoC code (self-contained):

2. Cluster PoC (Knative host injection, primary / practical bypass):

• Preconditions:
  • Kubernetes test cluster with Knative Serving.
  • Traefik configured with Knative provider.
• Apply manifest:
  • kubectl apply -f - <<'YAML'

• (If API version mismatch, adjust between networking.internal.knative.dev/v1alpha1 and networking.knative.dev/v1alpha1.)
• Verify:
  • Check Traefik router rule contains: (Host(tenant.example.com) || Host(attacker.com)) && PathPrefix(/).
  • Request with Host: attacker.com returns backend 200.
  • This demonstrates host restriction bypass in practice.

3. Cluster PoC (Knative header injection, confirms rule-syntax injection):

• Apply:
  • kubectl apply -f - <<'YAML'

• Verify:
  • Inspect generated Traefik dynamic router rule (API/dashboard/logs).
  • Confirm injected fragment || Host(attacker.com) is present.
  • Send request with Host: attacker.com and no expected tenant header (expected: 404 for this payload shape, because leading Host(tenant) still applies).
  • Send request with Host: tenant.example.com and X-Poc: x (expected: 200 from backend).

4. Optional Ingress PoC (scope check):

• Apply:
  • kubectl apply -f - <<'YAML'

• Expected in most clusters: API server rejects this payload because Ingress host must satisfy DNS-1123.
• Keep this step only as a negative control to demonstrate the distinction between native Ingress validation and Knative CRD behavior.

Validation executed in this report:

• Local deterministic PoC executed with go run and output matched expected injected rule.
• Live cluster test executed on local kind cluster (kind-traefik-poc) with Traefik v3.6.10 and Knative Serving CRDs.
• Annotation key confirmed in this environment: networking.knative.dev/ingress.class (dot). The hyphen variant was not used by the successful processing path.
• Traefik API/logs confirmed generated routers included injected expressions.
• Live HTTP request with Host: attacker.com reached backend (200) for Knative host-injection payload.

Impact

• Vulnerability type: Rule injection / authorization bypass at routing layer.
• Primary impact: Bypass of intended routing predicates (host/header/path), enabling unauthorized routing to protected services.
• Who is impacted: Primarily deployments using Traefik Knative provider where untrusted or semi-trusted actors can create/update Knative Ingress resources (typical in multi-tenant clusters, shared namespaces, or weak admission controls). Standard Kubernetes Ingress host injection is usually blocked by API validation.
• Security consequences: Cross-tenant traffic access, internal service exposure, policy bypass, and potential chaining with app-level vulnerabilities.