| name | security-review |
| description | Security audit of uncommitted changes. Scans for OWASP top 10, injection, exposed secrets, auth bypasses, and unsafe patterns. |
| argument-hint | [path] |
| context | fork |
Security Review
Dedicated security audit of code changes. Focuses exclusively on security vulnerabilities — not code quality, style, or correctness (those are /review's job).
Usage
/security-review
/security-review backend/src/
Behavior
1. Gather the Diff
git diff HEAD
If arguments provided, scope the diff:
git diff HEAD -- <path>
If no uncommitted changes, say so and stop.
2. Read Full Files for Context
For each changed file, read the full file (not just diff hunks). Security issues often span multiple functions.
3. Scan for Vulnerability Categories
A. Injection
- SQL injection: Raw string interpolation in queries, missing parameterized queries
- NoSQL injection: Unsanitized user input in MongoDB/Strapi filters
- Command injection: User input passed to shell execution without sanitization
- XSS: Unsanitized user content rendered as HTML, missing output encoding
- Path traversal: User input in file paths without sanitization
B. Authentication & Authorization
- Auth bypass: Missing auth checks on new routes/endpoints
- Privilege escalation: User can access/modify resources they shouldn't own
- JWT/token issues: Hardcoded secrets, missing expiration, weak algorithms
- IDOR: Direct object references without ownership validation
C. Data Exposure
- Secret leaks: API keys, passwords, tokens hardcoded in source (not env vars)
- Sensitive data in logs: PII, credentials, tokens logged to console/files
- Verbose errors: Stack traces or internal paths exposed to clients
- Over-fetching: API returns more fields than the client needs
D. Configuration
- CORS misconfiguration: Overly permissive origins, credentials with wildcards
- Missing rate limiting: New endpoints without rate limiting on auth/sensitive routes
- Insecure defaults: Debug mode in production, permissive CSP, missing security headers
- Dependency issues: Known vulnerable package versions in changed package.json
E. Unsafe Code Patterns
- Dynamic code execution: Functions that interpret strings as code with user-controlled input
- Unsafe HTML rendering: Setting inner HTML content without sanitization
- Prototype pollution: Object merging/spreading with unsanitized user objects
- Race conditions: TOCTOU bugs in auth checks, file operations
- Unvalidated redirects: User-controlled redirect URLs without allowlist
4. Assign Severity
- [CRITICAL] — Exploitable now. Auth bypass, injection, secret leak. Must fix immediately.
- [HIGH] — Likely exploitable with effort. IDOR, XSS, privilege escalation.
- [MEDIUM] — Needs specific conditions. Race condition, verbose errors, missing rate limit.
- [LOW] — Defense in depth. Missing security header, overly broad CORS in dev.
5. Self-Check
Before reporting:
- Is this actually exploitable, or am I speculating about a theoretical attack?
- Is user input actually reaching this code path?
- Does the framework already protect against this (e.g., React auto-escapes, Strapi has built-in auth)?
- Is this a dev-only configuration that won't reach production?
Drop findings below 70% confidence.
6. Fix Every Finding
Fix every finding you reported — CRITICAL, HIGH, MEDIUM, AND LOW. A clean security review leaves zero open findings, not just zero CRITICAL/HIGH.
For each finding:
- Read the full file for context.
- Apply the fix (parameterize the query, add the missing auth/ownership check, move the secret to an env var, narrow the over-fetch, restrict CORS, add the rate limit, etc.).
- Add a regression test where one is meaningful — e.g. a test asserting the endpoint returns 403 for a non-owner, or that a query rejects an injection payload. Follow the project's existing test conventions.
- Briefly note what changed and what test was added.
Only leave a finding unfixed if (a) it's a verified false positive — say so and why — or (b) the fix needs a threat-model decision (e.g. "should this role be allowed here at all?") that only the user can make — surface it explicitly and ask. Never defer a real finding silently.
Run the test suite after fixes to confirm nothing broke.
7. Output
## Security Review
Scanned N files with uncommitted changes.
### Findings
1. **[CRITICAL] Brief title** — `file:line`
Description of the vulnerability and how it could be exploited.
2. **[HIGH] Brief title** — `file:line`
Description.
### Verdict: SECURE / NEEDS INPUT
Each finding line should end with Fixed: a one-line description of the fix (and the regression test, if added).
Verdicts:
- SECURE: Zero open findings at any severity (CRITICAL through LOW) — everything found was fixed.
- NEEDS INPUT: A finding remains only because it needs a user threat-model decision or was flagged as a false positive. List exactly which, and why.
If zero findings, report "No security issues found" with verdict SECURE.
Rules
- Security only. Don't flag code quality, style, or correctness issues — that's
/review's domain.
- No false positives. A wrong security finding causes alert fatigue. When in doubt, don't flag.
- Find AND fix. Every real finding (CRITICAL→LOW) gets fixed, not just reported. A clean review leaves zero open findings.
- Brief. One paragraph max per finding: describe the attack vector, then state the fix in one line.
- Respect frameworks. Don't flag things the framework already handles (React escaping, Strapi auth middleware).
- Read context. A SQL query in a migration script is not injection. A hardcoded string in a test is not a secret leak.
