installed-skill-auditor
Weekly audit of all installed third-party and community skills for malicious patterns, stale credentials, and drift from last-known-good state.
git clone --depth 1 https://github.com/ArchieIndian/openclaw-superpowers /tmp/installed-skill-auditor && cp -r /tmp/installed-skill-auditor/skills/openclaw-native/installed-skill-auditor ~/.claude/skills/installed-skill-auditorSKILL.md
# Installed Skill Auditor
## What it does
`skill-vetting` scans before install. `installed-skill-auditor` scans after — continuously.
Skills can be modified after installation. A community skill that was safe on Monday can be compromised by Tuesday if the source repo is pushed to and your agent auto-pulls. This skill runs weekly to catch post-install drift: injected payloads, hardcoded credentials, and pattern changes that weren't there at install time.
It maintains a content hash of every skill file at the time it was first audited. On each weekly run it re-hashes and flags anything that changed unexpectedly.
## When to invoke
- Automatically, every Monday at 9am (cron)
- Manually after any `git pull` that touches skill directories
- After any agent action that writes to the skills tree
## Audit checks
| Check | What it detects |
|---|---|
| INJECTION | Instruction-override patterns in SKILL.md prose |
| CREDENTIAL | Hardcoded tokens, API keys, or secrets in any file |
| EXFILTRATION | URLs + data-sending patterns suggesting exfil |
| DRIFT | File content changed since last known-good baseline |
| ORPHAN | Skill directory present but not in install manifest |
Severity: CRITICAL (INJECTION, EXFILTRATION) · HIGH (CREDENTIAL) · MEDIUM (DRIFT, ORPHAN)
## Output
```
Installed Skill Audit — 2026-03-16
────────────────────────────────────────────
32 skills audited | 0 CRITICAL | 1 HIGH | 2 MEDIUM
HIGH community/my-custom-skill — CREDENTIAL
Hardcoded token pattern detected in run.py (line 14)
MEDIUM community/expense-tracker — DRIFT
SKILL.md hash changed since 2026-03-10 baseline
Run: python3 audit.py --diff expense-tracker
```
## How to use
```
python3 audit.py --scan # Full audit pass
python3 audit.py --scan --critical-only # CRITICAL findings only
python3 audit.py --baseline # Record current state as trusted
python3 audit.py --diff <skill-name> # Show changed lines since baseline
python3 audit.py --resolve <skill-name> # Mark finding resolved after review
python3 audit.py --status # Summary of last run
python3 audit.py --format json # Machine-readable output
```
## Procedure
**Step 1 — Review the report**
The cron run generates a report automatically. Open it via `--status` or check state. Any CRITICAL finding requires immediate action.
**Step 2 — Triage by severity**
- **CRITICAL**: Do not run the skill. Inspect the file, remove or quarantine the skill.
- **HIGH**: Rotate the exposed credential immediately; investigate how it got there.
- **MEDIUM (DRIFT)**: Use `--diff` to see what changed. If the change is expected (you updated the skill), run `--baseline` to accept it. If unexpected, treat as CRITICAL.
- **MEDIUM (ORPHAN)**: A skill directory exists with no install record. Either re-install through the vetting process or remove the directory.
**Step 3 — Resolve or escalate**
Run `--resolve <skill-name>` after reviewing a finding. This marks it acknowledged in state. Unresolved CRITICAL findings are surfaced again on next cron run.
**Step 4 — Update baseline after intentional changes**
When you intentionally update a skill (e.g., upgrading to a new version), run `--baseline` so future drift detection has an accurate reference point.
## State
Results and content hashes stored in `~/.openclaw/skill-state/installed-skill-auditor/state.yaml`.
Fields: `last_audit_at`, `baselines` (hash map), `findings`, `audit_history`.Syncs agent daily memory and MEMORY.md to an Obsidian vault so notes are human-browsable. Use nightly or on demand.
Structured ideation before any implementation. Use when starting any non-trivial task.
Scaffolds and validates new superpowers skills. Use when creating a new skill for this repository.
Executes plans task-by-task with verification. Use when implementing a plan.
Triggers a secondary verification pass for any agent output containing factual claims, numbers, dates, or named entities before the output is acted on
Crawls a new codebase to infer stack, conventions, and key invariants, then generates a PROJECT.md context file for the agent
Handles PR review feedback by fetching comments, grouping issues, fixing one group at a time, and verifying before replies.
Detects skill name shadowing and description-overlap conflicts that cause OpenClaw to trigger the wrong skill or silently ignore one when two skills compete for the same intent.