| name | security-review |
| description | Audit code changes for security vulnerabilities before merging. Use when reviewing a diff/PR, hardening code, or the task mentions security, injection, XSS, SSRF, secrets, auth, deserialization, path traversal, or "is this safe?". Provides a concrete checklist and reporting format; never auto-fixes silently. |
| tools | ["Bash","Read","Grep","Glob"] |
Security Review
A focused checklist for catching real vulnerabilities in a change set. Apply it
to the diff first, then widen only if a finding points elsewhere. Report
findings; do not rewrite code unless explicitly asked.
Threat checklist
- Injection — untrusted input flowing into SQL, shell, OS commands, eval,
template engines, or LDAP. Look for string concatenation instead of
parameterized queries/argument arrays.
- XSS — user data rendered into HTML/JS/DOM without escaping. Check
innerHTML, dangerouslySetInnerHTML, unescaped template interpolation.
- AuthN / AuthZ — missing authentication, missing ownership/permission
checks, IDOR, privilege escalation, trusting client-supplied roles.
- Secrets — hardcoded API keys, tokens, passwords, private keys, or
connection strings; secrets logged or committed.
- Path traversal / file access — user-controlled paths joined without
normalization/allow-listing (
../ escapes, symlink following).
- SSRF — server-side requests to a user-supplied URL/host without
allow-listing; can reach internal metadata endpoints.
- Deserialization / parsing — untrusted data fed to unsafe deserializers
(pickle, native YAML loaders, Java/PHP object deserialization).
- Crypto — weak algorithms (MD5/SHA1 for passwords), hardcoded IVs,
ECB mode, missing TLS verification, predictable randomness for tokens.
- Sensitive data exposure — PII/secrets in logs, error messages, or API
responses; verbose stack traces leaked to clients.
- Dependencies — newly added packages: are they reputable, pinned, and free
of known CVEs? Avoid typosquats.
- Resource & DoS — unbounded loops, unbounded request/body sizes, missing
timeouts, regex catastrophic backtracking (ReDoS).
- Race conditions / TOCTOU — check-then-act on files, balances, or auth
state without locking.
Method
- Identify the trust boundary: where does untrusted input enter, and where
does it reach a sink (DB, shell, filesystem, network, HTML)?
- Trace each tainted value from source to sink.
- Prefer allow-lists over deny-lists; prefer library-provided escaping and
parameterization over manual sanitization.
- Only flag issues you can justify with a concrete exploit path.
Report format
For each finding:
[severity: high|medium|low] <title>
location: path/to/file.ext:LINE
issue: <what is wrong and the input that triggers it>
impact: <what an attacker gains>
fix: <the minimal concrete remediation>
Order findings by severity. If you reviewed the change and found nothing
actionable, say so explicitly.
