autoclaw-security - SKILL.md Agent Skill

name: autoclaw-security description: Security review for PANaCEa. Audit secrets, auth, dependencies, external skills, and risky operations. Use before migrations, deploys, auth changes, or external installs. mode: security

Security Mode — Risk Reduction

Purpose

Reduce blast radius. Never leak secrets. Verify auth boundaries.

When to Use

Before running external skills or scripts
Before installing dependencies
Before schema migrations
Before deployment changes
Before auth middleware changes
Before any command with elevated permissions
When reviewing sub-agent output for security issues

Audit Checklist

Secrets

No API keys, tokens, or secrets in code?
No DB URLs exposed in client code?
No env vars leaking to browser?
.env in .gitignore?

Auth

User identity derived from auth/session, not client input?
Ownership checked before data mutation?
Auth not disabled to make feature work?
CORS not loosened without reason?
Rate limiting present for public endpoints?

Code

No process.env in Edge functions?
safePrismaDisconnect in all Edge finally blocks?
Input validated (Zod or manual)?
SQL injection vectors checked?
XSS vectors in user-generated content?

Dependencies

New dependency inspected for: source, license, maintenance, advisories?
No abandoned or suspicious packages?
Bundle impact assessed?

External Skills/Scripts

Source verified?
All commands inspected?
Network calls understood?
No obfuscated code?
No secret exfiltration?
Sandboxed where possible?

Output Format

## Security Review: {target}
**Risk Level:** 🟢 Low / 🟡 Medium / 🔴 High

### Issues
1. 🔴 {critical issue} — {file:line} — {fix}
2. 🟡 {warning} — {file:line} — {mitigation}

### Blast Radius
**Files affected:** {count}
**Data affected:** {description}
**Rollback:** {how to undo}

### Verdict
✅ Safe to proceed / ⚠️ Proceed with mitigations / ❌ Blocked

Logged to .autoclaw/security-log.md

Coordination

Triggered by: Orchestrator (before risky ops), Reviewer (security concerns found), Architect (migration/deploy designs)
Hands off to: Architect (risk report for design adjustment), Orchestrator (go/no-go signal)
Critical subsystems: RISK-001 (FSRS), RISK-002 (Session), RISK-003 (Edge Auth) — see docs/autoclaw/coordination/risk_register.md

Pre-Flight

# Check risk register for known subsystem vulnerabilities
grep -A10 "RISK-00[1-3]" docs/autoclaw/coordination/risk_register.md
# Verify secrets hygiene
rg "process\.env" functions/api/ 2>/dev/null || echo "Edge hygiene: clean"
# Check .gitignore covers secrets
grep "\.env" .gitignore

Common Pitfalls

Skipping blast radius analysis: Always assess what else the change affects
Trusting external code: Sub-agent and external script output needs full security review
Auth bypass in tests: Tests that skip auth don't prove production safety