Security Review

Security Review

Audit changed files for security vulnerabilities, focusing on the OWASP Top 10 and issues specific to the project's stack.

When running locally as a forked subagent, the main session does not see any files you read or any reasoning you do — only the final report you return. When running in CI (e.g. via claude-code-action), the workflow takes the report and turns it into GitHub PR review comments. Either way, take your time, read every changed file completely, and produce a thorough, actionable report. The consumer of this report uses it as a worklist, so it must be complete and self-contained.

Scope

Determine the diff to review:

1. Run git diff main...HEAD --name-only to get files changed on this branch vs main.
2. If that fails (no main, detached worktree, etc.), fall back to git diff HEAD --name-only for uncommitted changes, then git diff --cached --name-only for staged files.
3. If no diff is available, ask the user which files to review.

Read every changed file completely before starting the review. Read CLAUDE.md first to understand the project's stack and any subsystems with security-sensitive surface area (auth, real-time, payments, file uploads).

What to Look For

Injection & Input Handling

• SQL injection: Raw SQL with string interpolation instead of parameterized queries. Check for any template literals that build SQL, and verify that database libraries are being used in their parameterized form.
• Command injection: User input passed to shell commands (Bun.$, child_process, subprocess, os.system) without sanitization.
• XSS: User-controlled data rendered as dangerouslySetInnerHTML, or reflected into HTML/JS without escaping. Check contentEditable fields that accept pasted HTML.
• Path traversal: User input used in file paths without validation. Check for .. traversal.
• Prototype pollution: Object.assign or spread on user-controlled objects without allowlisting keys.

Authentication & Authorization

• Missing auth checks: Endpoints that read/write data without verifying the caller's identity or org membership.
• IDOR (Insecure Direct Object Reference): Endpoints that accept an ID parameter and return/modify the resource without verifying the caller has access. Particularly dangerous when URL params (like a slug) aren't validated against the actual resource ownership.
• Privilege escalation: Actions that should be restricted (delete, move, admin operations) but aren't gated on role/permission.

Data Exposure

• Over-fetching: API responses that include more data than the client needs (e.g., internal IDs, secrets, full document state when only a title is needed).
• Error leakage: Stack traces, SQL errors, or internal paths exposed in error responses.
• Sensitive data in logs: Passwords, tokens, or PII logged to console.

Real-time / WebSocket Security