security-audit

A hostile-by-design, fail-closed audit workflow for codebases and OpenClaw/ClawHub skills.

It does not try to answer “does this skill work?”. It tries to answer: “can this skill betray the system?”

What it checks (high level)

This skill’s scripts combine multiple layers:

• Secrets / credential leakage: trufflehog
• Static analysis: semgrep (auto rules)
• Hostile repo audit (custom): prompt-injection signals, persistence mechanisms, suspicious artifacts, dependency hygiene

If any layer fails, the overall audit is FAIL.

Run an audit (JSON)

From this skill folder (use bash so it works even if executable bits were not preserved by a zip download):

bash scripts/run_audit_json.sh <path>

Example:

bash scripts/run_audit_json.sh . > /tmp/audit.json
jq '.ok, .tools' /tmp/audit.json

Security levels (user configurable)

Set the strictness level (default: standard):

OPENCLAW_AUDIT_LEVEL=standard bash scripts/run_audit_json.sh <path>
OPENCLAW_AUDIT_LEVEL=strict   bash scripts/run_audit_json.sh <path>
OPENCLAW_AUDIT_LEVEL=paranoid bash scripts/run_audit_json.sh <path>

• standard: pragmatic strict defaults (lockfiles required; install hooks/persistence/prompt-injection signals fail)
• strict: more patterns become hard FAIL (e.g. minified/obfuscation artifacts)
• paranoid: no "best-effort" hashing failures; more fail-closed behavior

Manifest requirement (for zero-trust install workflows)

For strict/quarantine workflows, require a machine-readable intent/permissions manifest at repo root:

• openclaw-skill.json

If a repo/skill does not provide this manifest, the hostile audit should treat it as FAIL.

See: docs/OPENCLAWSKILLMANIFEST_SCHEMA.md.

Optional: execution sandbox (Docker)

Docker is optional here. This skill can be used for static auditing without Docker.

If you want to execute any generated/untrusted code, run it in a separate sandbox workflow (recommended).

Files

• scripts/runauditjson.sh — main JSON audit runner
• scripts/hostile_audit.py — prompt-injection/persistence/dependency hygiene scanner
• scripts/security_audit.sh — convenience wrapper (always returns JSON, never non-zero)
• openclaw-skill.json — machine-readable intent/permissions manifest