Summary

The Kubernetes admission webhook handler reads the entire request body using io.ReadAll(r.Body) without any size limit. Any client that can reach the webhook port within the cluster can send a multi-GB payload, causing the skipper process to exhaust memory and be OOM-killed. This disrupts all Kubernetes admission control, potentially blocking all pod creation and updates.

Vulnerable Code

// dataclients/kubernetes/admission/admission.go:76
body, err := io.ReadAll(r.Body)  // <-- NO SIZE LIMIT
if err != nil {
    log.Errorf("Failed to read request: %v", err)
    w.WriteHeader(http.StatusInternalServerError)
    invalidRequests.WithLabelValues(admitterName).Inc()
    return
}

var review admissionReview
err = json.Unmarshal(body, &review)

For comparison, the OPA filter has a body size limit:

// filters/openpolicyagent/openpolicyagent.go:68-70
const DefaultMaxRequestBodySize = 1 << 20 // 1MB

// OPA uses a bufferedBodyReader with size limits

Attack Path

1. Attacker identifies the admission webhook endpoint (default: :9443/admission or configured path)
2. Attacker sends: POST /admission HTTP/1.1, Content-Type: application/json with a multi-GB request body
3. io.ReadAll(r.Body) allocates unbounded memory for the entire body
4. Skipper process is OOM-killed by the Kubernetes kubelet

Permission Boundary Analysis

• Attacker: Any client with network access to the admission webhook port within the Kubernetes cluster
• Boundary crossed: Memory safety — unbounded allocation from attacker-controlled input
• Preconditions: Admission webhook endpoint must be network-reachable (default Kubernetes deployment exposes it within cluster network)
• Comparison: OPA filter has DefaultMaxRequestBodySize (1MB) and semaphore-based memory limit; admission handler has neither

Evidence

Tests

• dataclients/kubernetes/admission/admission_test.go exists but does not test body size limits

Impact

The admission webhook handler reads the entire request body using io.ReadAll(r.Body) without a size limit. An attacker with in-cluster network access and a valid Kubernetes client certificate can send a multi-GB payload to the webhook endpoint, causing the skipper process to exhaust memory and be OOM-killed. This disrupts admission control for Ingress and RouteGroup resources until the process is automatically restarted by the kubelet.

Scope of impact: Ingress and RouteGroup admission only — not pod creation or other admission controllers.

Recovery: Kubernetes automatically restarts the OOM-killed process, limiting downtime.

Prerequisites: (1) In-cluster network access to the webhook port, (2) valid Kubernetes client certificate.

Mitigation

1. Add http.MaxBytesReader or equivalent body size limit before io.ReadAll
2. Follow the OPA filter pattern: define DefaultMaxRequestBodySize and use a buffered reader with size limits
3. Add a configurable --admission-max-body-size flag

References

• GHSA-cwxq-rc9x-2jvv
• https://github.com/zalando/skipper/releases/tag/v0.26.22