SuperAPI AI Directory

security-review

Security & auditing

View repository

Updated 10-16 07:25

About this skill

Security Review

Review the current change like a security-minded senior. Prefer concrete findings over generic advice.

Scope

Focus on the diff and directly related call paths. Do not audit the whole repo unless asked.

Checklist

Auth & access

  • Every sensitive action checks authentication and authorization
  • IDs from the client are not trusted for ownership (userId in body ≠ proof)
  • Admin/debug routes are gated; no "temporary" open endpoints

Input & data

  • User input is validated at the boundary (type, length, allowlist)
  • SQL/NoSQL/search queries are parameterized — no string concat
  • HTML/Markdown/user content is escaped or sanitized where rendered
  • File uploads: size limits, type allowlist, stored outside web root, randomized names
  • Path joins cannot escape intended directories (../)

Secrets & config

  • No secrets in source, logs, client bundles, or example env files with real values
  • New env vars documented; defaults are safe for production
  • Tokens/passwords not written to analytics or error trackers

Network & server

  • Outbound fetches do not accept raw user URLs without allowlisting (SSRF)
  • CORS is least-privilege; credentials only when required
  • Cookies: HttpOnly, Secure, SameSite where applicable

Dangerous APIs

  • No eval, dangerouslySetInnerHTML, shell exec with unsanitized input
  • Deserialization of untrusted data is avoided or hardened
  • Crypto/auth libraries are standard — no home-rolled JWT/crypto

Output format

Group findings:

Severity Rule Where Why it matters Fix
Critical / High / Medium / Low / Note short name file:line or symbol 1 sentence concrete change

End with:

  • Blockers — must fix before merge
  • Safe to ship with follow-ups — optional

If no issues: say so explicitly and list what you checked.