Claw Permission Firewall

Evaluates agent actions for security risks, enforcing least-privilege policies with allow, deny, or confirmation decisions and secret redaction.

Runtime least-privilege firewall for agent/skill actions. It evaluates a requested action and returns one of:

• ALLOW (safe to execute)
• DENY (blocked by policy)
• NEED_CONFIRMATION (risky; require explicit confirmation)

It also returns a sanitizedAction with secrets redacted, plus a structured audit record.

This is not a gateway hardening tool. It complements gateway security scanners by enforcing per-action policy at runtime.

---

What it protects against

• Exfiltration to unknown domains
• Prompt-injection “send secrets” attempts (secret detection + redaction)
• Reading sensitive local files (~/.ssh, ~/.aws, .env, etc.)
• Unsafe execution patterns (rm -rf, curl | sh, etc.)

---

Inputs

Provide an action object to evaluate:

json

{
  "traceId": "optional-uuid",
  "caller": { "skillName": "SomeSkill", "skillVersion": "1.2.0" },
  "action": {
    "type": "http_request | file_read | file_write | exec",
    "method": "GET|POST|PUT|DELETE",
    "url": "https://api.github.com/...",
    "headers": { "authorization": "Bearer ..." },
    "body": "...",
    "path": "./reports/out.json",
    "command": "rm -rf /"
  },
  "context": {
    "workspaceRoot": "/workspace",
    "mode": "strict | balanced | permissive",
    "confirmed": false
  }
}

---

Outputs

json

{
  "decision": "ALLOW | DENY | NEED_CONFIRMATION",
  "riskScore": 0.42,
  "reasons": [{"ruleId":"...","message":"..."}],
  "sanitizedAction": { "...": "..." },
  "confirmation": { "required": true, "prompt": "..." },
  "audit": { "traceId":"...", "policyVersion":"...", "actionFingerprint":"..." }
}

---

Default policy behavior (v1)

• Exec disabled by default
• HTTP requires TLS
• Denylist blocks common exfil hosts (pastebins, raw script hosts)
• File access is jailed to workspaceRoot
• Always redacts Authorization, Cookie, X-API-Key, and common token patterns

---

Recommended usage pattern

1. Your skill creates an action object.
2. Call this skill to evaluate it.
3. If ALLOW → execute sanitizedAction.
4. If NEED_CONFIRMATION → ask user and re-run with context.confirmed=true.
5. If DENY → stop and show the reasons.

---

Files

• policy.yaml contains the policy (edit for your environment).