gke-cred-audit

Defensive credential-exposure auditor for GKE. Deployed in-cluster as a single Pod (Deployment + Service); fronted by Natoma at /mcp.

The server uses its own ServiceAccount to:

• Inventory its own GCE/GKE workload-identity exposure (project, instance, SA scopes/tokens).
• Decode its own projected ServiceAccount token (claims only) and walk its mounted secret files.
• Inventory the pods + ConfigMaps in its namespace (and, opt-in, secret metadata) so Natoma can ask follow-up questions about workloads that share the namespace.
• Compute its own RBAC posture via SelfSubjectRulesReview and a curated SelfSubjectAccessReview matrix.

Surfaces

• REST with an OpenAPI 3.1 spec at /openapi.json (Swagger UI at /docs, ReDoc at /redoc).
• MCP server over Streamable HTTP, mounted at /mcp (per MCP spec rev 2025-03-26).

Both surfaces share the same redacted pydantic models. Raw bearer tokens, JWT signatures, PEM private keys, and Secret values are never returned. The break-glass RAW_REVEAL flag exists for debugging individual tokens; it never extends to namespace Secret values.

Deploying via Natoma

Natoma is the auth boundary for /mcp. The server itself has no native auth; this is intentional and surfaced in /version, audit://server-info, and the OpenAPI spec via x-auth=none (gateway-managed).

Dockerfile path (recommended)

This repo has a root Dockerfile. Point Natoma at this repo and it will build and run the container. The server reads $PORT (default 8080) and binds 0.0.0.0. uvicorn is configured with --proxy-headers --forwarded-allow-ips="*" so X-Forwarded-* from the gateway is honored.

Direct cluster deployment

kubectl apply -f manifests/rbac.yaml
kubectl apply -f manifests/deployment.yaml
kubectl apply -f manifests/service.yaml
kubectl apply -f manifests/networkpolicy.yaml

Then point Natoma's gateway at http://gke-cred-audit.<namespace>.svc:8080/mcp.

The bundled NetworkPolicy only permits ingress from namespaces labeled natoma-gateway: "true"; adjust to match your installation. Egress is restricted to the Kubernetes API server and the GCE metadata IP.

Granting secret enumeration

AUDIT_ENABLE_SECRET_LISTING=true opts in to namespace Secret enumeration. Even when enabled, the server returns metadata only -- name, type, key names, per-key size, per-key SHA-256 prefix. Values are never returned. A property-based test (tests/test_redaction_invariant.py) runs against mocked Secrets containing real-looking base64 data to enforce this.

The trade-off: granting secrets:list to the audit ServiceAccount makes the audit pod a high-trust target. Default is OFF. When enabling, prefer:

• A dedicated namespace for the audit Deployment, with no other workloads.
• Tightening the bundled Role with resourceNames: [...] to specific Secret names.
• A NS-SECRETS-LIST-GRANTED finding will appear in /findings to make this visible.

Running locally

pip install -e '.[dev]'
gke-cred-audit --bind 127.0.0.1 --port 8080

Then:

• curl http://127.0.0.1:8080/openapi.json -- OpenAPI 3.1 document
• curl http://127.0.0.1:8080/findings?severity=HIGH -- JSON findings
• curl http://127.0.0.1:8080/server-info -- capability manifest
• MCP clients connect to http://127.0.0.1:8080/mcp

What's intentionally NOT in scope

• Cross-namespace reads (would require ClusterRole; not used).
• Returning Secret data/stringData under any flag.
• Native auth on the server (Natoma owns auth).