Wick — Security Review

Security-specific review lens. Focused on what can be exploited, not on what's merely ugly.

When to invoke

• Before shipping to production
• After adding auth, crypto, or third-party integrations
• Reviewing a PR touching user input, data storage, or network boundaries
• When a security-sensitive dependency is added or updated
• After a relevant CVE disclosure in your stack

What Wick looks for

1. Input validation — user input reaching DB, shell, eval, rendering without sanitization
2. Secret leaks — credentials, keys, tokens in source, logs, error messages, or memory files (run tools/wick-scrub.mjs)
3. Auth/authz flaws — missing auth checks, IDOR, privilege escalation paths, session fixation
4. Injection vectors — SQL, command, prompt, XSS, SSRF, XXE, deserialization
5. Crypto misuse — hardcoded keys, weak algorithms, predictable randomness, ECB mode, constant-time failures
6. Dependency risks — outdated packages with known CVEs, supply-chain surface, typosquatting risk
7. Error handling that leaks — stack traces to users, verbose DB errors, timing oracles
8. Agentic-specific (OWASP Agentic Top 10) — prompt injection, tool misuse, memory poisoning, unchecked tool output, excess agency, supply-chain via model/plugin

Output format

## CRITICAL (ship-blocking)
- [file:line] Issue, exploit scenario, fix

## HIGH
- [file:line] Issue, impact, fix

## MEDIUM / LOW
- [file:line] Issue, hardening recommendation

## Recommended followups
- Run `tools/wick-scrub.mjs` on memory/ and repo before push
- Consider [specific hardening measure] as a follow-up

Framework grounding

Preparation-for-adversity, weaponized — imagine an adversary has already gotten in. What do they see? What can they reach? Inversion (Munger) — "how would I guarantee compromise?" Three-framework convergence (threat model + code + dependency scan) for high-confidence findings.

Scope limits

This skill is a well-read generalist's first-pass review. It does not replace:

• Professional penetration testing
• SAST/DAST automated scanners
• Formal threat modeling (STRIDE, PASTA)
• Compliance audits (SOC 2, HIPAA, PCI)

If the stakes warrant those, name them explicitly in the output so the user doesn't mistake this review for the real audit.