Security Review

<!--

Source: anthropics/claude-code-security-review (official)

File: https://github.com/anthropics/claude-code-security-review/blob/main/.claude/commands/security-review.md

Used by: reviewer agent

-->

Security Review

When to trigger

• After writing code touching auth, payments, API endpoints, file I/O, or user input
• Before commit on security-sensitive changes
• When user asks /security-review

Scanning mode

Diff-aware: only analyze changed files in the current PR/commit.

Focus on real vulnerabilities — filter out false positives aggressively.

What to check (OWASP Top 10 mapping)

1. Injection

• SQL injection → parameterized queries?
• NoSQL injection → $where clauses, user-controlled operators?
• Command injection → exec() with user input?
• Prompt injection (AI-specific) → user input flowing into system prompts?

2. Broken authentication

• Missing auth checks on protected routes
• Weak password hashing (bcrypt/argon2 required)
• JWT misuse — missing signature verification, weak secrets, none algorithm accepted
• Session fixation, missing session rotation on login

3. Sensitive data exposure

• Hardcoded secrets in code
• API keys in client bundles (import in "use client" components)
• PII in logs
• Missing HTTPS enforcement
• Secrets in error messages

4. XML/XXE

• XML parsers with external entity processing enabled

5. Broken access control

• Missing authorization (authenticated but not authorized)
• IDOR (accessing other users' data by changing an ID)
• Privilege escalation paths

6. Security misconfiguration

• Default credentials
• Debug endpoints exposed in prod
• Missing security headers (CSP, HSTS, X-Frame-Options)

7. XSS

• dangerouslySetInnerHTML with user input
• Rendering LLM output without sanitization
• URL parameters reflected without escaping

8. Insecure deserialization

• JSON.parse with user-controlled content (usually fine in JS, but watch for prototype pollution)

9. Using components with known vulnerabilities

• Run npm audit mentally — are any flagged deps in critical paths?

10. Insufficient logging

• Security events unlogged (failed logins, permission denials)
• Overlogging (logging PII, tokens, passwords)

AI-specific checks

• Webhooks: signature verification present? (Stripe, GitHub, etc.)
• Rate limiting: on public endpoints? On AI endpoints especially?
• Input length: max length on LLM inputs?
• Output validation: Zod schema on LLM responses before rendering?
• CORS / CSRF: properly configured for auth endpoints?

Severity (same as code-review)

• CRITICAL — exposed secret, missing auth, unverified webhook, SQL injection, missing rate limit on AI endpoint
• HIGH — missing input validation, weak crypto
• MEDIUM — logging gaps, overly verbose errors
• LOW — defense-in-depth suggestions

Output format

## CRITICAL
- [`<file>:<line>`] <vulnerability> (OWASP: <category>)
  - Impact: <what an attacker could do>
  - Fix: <specific mitigation>

## HIGH / MEDIUM / LOW
- ...

## Verdict
PASS | BLOCK (CRITICAL count: N)

Rules

• Never hand-wave. "This might be a risk" is not enough — say if it is or isn't, with evidence.
• Every finding: severity, OWASP category, file + line, attacker scenario, specific fix.
• Rotate suspected exposed secrets immediately and flag at top of report.
• False positives are worse than no review — if you're not sure, don't flag it.
• AI endpoints get extra scrutiny — token limits, output sanitization, rate limits.