| name | security-audit |
| description | Use when auditing a codebase before it goes public-facing, after a lab build, or on any periodic security pass. Checks secrets in code and history (including Claude Code transcripts), authorization and RLS, input validation, dependency red flags, and exposed endpoints. |
Security Audit
Run this in a FRESH session. An auditor that helped write the code shares its blind spots. Zero context is the feature: /clear first, or open a new session, then invoke this skill.
Work through all five fronts. For each finding report: severity (critical/high/medium/low), file and line, the attack it enables, and the fix. No hand-waving; every finding gets a concrete fix.
1. Secrets in code and history
- Grep the working tree for key patterns:
sk-, sk_live, AKIA, ghp_, -----BEGIN, password\s*=, token\s*= with literal values.
- Check git history, not just HEAD:
git log -p | grep for the same patterns. A key deleted in a later commit is still leaked.
- Check
.env is gitignored and no .env* file is tracked.
- The one everyone misses: Claude Code transcripts. Every paste lands in
~/.claude session JSONL files forever. Search them: grep -rl "sk-\|sk_live\|AKIA" ~/.claude/projects/ 2>/dev/null. If keys appear, instruct the user to rotate them. Deleting the transcript does not unleak a key.
2. Authorization and RLS
- Every endpoint that reads or writes user data: where is the ownership check? "The frontend only shows your own data" is not a check.
- Supabase/Postgres: confirm RLS is enabled on every table with user data and policies are default-deny. List any table where RLS is off.
- Look for IDs accepted from the client and used in queries without verifying the session user owns them.
3. Input validation
- Every external input (request bodies, query params, webhooks, file uploads): is it validated at the boundary with a schema, or trusted?
- Raw string interpolation into SQL, shell commands, or HTML.
- File uploads: type, size, and path checks.
4. Dependencies
- Scan recently added dependencies for typosquats: names one edit away from popular packages (models hallucinate package names; attackers pre-register the typos). Verify each new package exists on the registry with real download numbers and a real repo.
- Flag packages with install scripts doing anything unusual.
- Run the ecosystem's audit command (
pnpm audit, pip-audit) and triage criticals.
5. Exposed endpoints
- Enumerate every route. For each: is it meant to be public? Debug routes, admin panels,
/dev/* prototypes, and webhook endpoints without signature verification all count as findings.
- Check CORS configuration and any wildcard origins.
- Confirm secrets are not shipped to the client bundle (search build output for key patterns).
Output
A findings table sorted by severity, then a "fix now" list of the top 3. Offer to fix the top finding immediately.
