Skill Audit

Audit locally installed agent skills for security/policy issues using the SkillLens CLI (skilllens scan, skilllens config). Use when asked to scan a skills directory (Codex/Claude) and produce a risk-focused audit report based on each skill's SKILL.md and bundled resources.

Skills Audit (SkillLens)

Install SkillLens

• One-off run: npx skilllens scan (or pnpm dlx skilllens scan)
• Global install: pnpm add -g skilllens

Quick start

• Run skilllens config to see configured scan roots and auditor CLI availability.
• Run skilllens scan to scan configured roots, or skilllens scan <path> to scan a specific directory.
• Re-run with --verbose to see raw auditor output and --force to ignore cached results.

Manual Fallback (when skilllens is unavailable)

If skilllens is not installed, perform a manual audit:

1. Inventory skills: find <path> -name "SKILL.md"
2. Scan each SKILL.md for the 6 risk categories:
   • Exfiltration: exec, subprocess, os.system, curl|sh, base64, wget — sends data out
   • Execution: eval, exec, shell injection, fetch-and-execute patterns
   • Persistence: cron, launch agent, shell profile, ~/.config mod
   • Privilege bypass: sudo, chmod 777, approval bypass, policy override
   • Prompt injection: "ignore previous", "always comply", "never reveal"
   • Overbroad trigger: vague description that fires on unrelated requests
3. Rate each skill: Safe (0 exec, 0 exfil) → Suspicious (some exec) → Unsafe (direct exfil or credential access)
4. Report: name, path, verdict, risk 0–100, evidence quotes, fix recommendations

Audit workflow

1. Define scope

   • Prefer a concrete target path (example: ~/.codex/skills) unless the user explicitly wants all configured roots.
   • If auditing a repo checkout containing skills, scan the parent folder that contains skill directories (example: skilllens scan ./skills).

2. Inventory skills with SkillLens

   • Run skilllens scan [path] [--auditor claude|codex].
   • Treat missing auditor CLIs or skipped statuses as “manual review required”, not “safe”.

3. Prioritize review order

   • Review any unsafe or suspicious verdicts first.
   • Next, review skills that request broad permissions (filesystem/network), run shell commands, or reference external downloads.

4. Manually review each skill’s contents

   • Read the skill’s SKILL.md and any referenced scripts/, references/, and assets/.
   • Do not execute bundled scripts by default; inspect first.

5. Evaluate risks (focus on realistic abuse)

   • Exfiltration: sending file contents, env vars, tokens, SSH keys, browser data, or configs to remote endpoints.
   • Execution: instructions to run arbitrary shell commands, curl | bash, eval, or to fetch-and-execute code.
   • Persistence: modifying shell profiles, launch agents, cron, editor configs, or skill install locations.
   • Privilege/approval bypass: instructions to ignore system policies, disable safety checks, or request escalated permissions unnecessarily.
   • Prompt injection: attempts to override higher-priority instructions (“ignore previous”, “always comply”, “never mention…”).
   • Overbroad triggers: vague descriptions that cause the skill to trigger on unrelated tasks.

6. Produce a report

   • For each skill, include: name, path, verdict (safe/suspicious/unsafe), risk (0–100), and bullet issues with concrete evidence (quote or filename).
   • Recommend fixes that reduce blast radius: narrow scope, remove dangerous defaults, add explicit confirmation gates, and document required permissions.

Command snippets

• Scan configured roots: skilllens scan
• Scan a specific folder: skilllens scan ~/.codex/skills
• Force a re-audit and show raw output: skilllens scan ~/.codex/skills --force --verbose