use-context-reviewer-security

Security Review

Detection (OWASP Top 10)

LLM01 covers apps that pass untrusted content to an LLM. The sink is the prompt itself: untrusted text concatenated without a data/instruction boundary, or a caller-supplied value interpolated into a system prompt. Constrain an LLM tool (e.g. a fetch_url tool) the same way as its non-LLM counterpart (A10 SSRF). The tool is the sink, not the prompt.

ID	Category	Pattern	Fix
A01	Broken Access Control	Missing auth, IDOR, path traversal	Auth middleware, ownership check
A02	Cryptographic Failures	password: 'plaintext'	bcrypt/argon2 hashing
A03	Injection	db.query(\SELECT...${id}`)`	Parameterized query, ORM
A03	Injection	exec(\ping ${host}`)`	Input validation, library instead
A03	XSS	dangerouslySetInnerHTML (static presence)	Default escaping, DOMPurify
A05	Security Misconfiguration	cors({ origin: '*' })	Explicit origin allowlist
A05	Security Misconfiguration	cookie: {} (no options)	secure, httpOnly, sameSite: 'strict'
A05	Security Misconfiguration	err.stack in error response without NODE_ENV guard	Generic message in prod, log internally
A09	Logging Failures	logger.info({ password })	Exclude sensitive fields
A10	SSRF	fetch(userInputUrl)	URL validation, allowlist
A07	Authentication Failures	No rate limit on auth endpoints (login, register, password-reset)	Rate limiter middleware on auth route group
A01	CSRF	No CSRF token verification on state-changing requests (POST/PUT/PATCH/DELETE)	Double Submit Cookie
A02	Timing Attack	=== comparison of tokens/signatures	Constant-time comparison (XOR all bytes, decide last)
A08	Prototype Pollution	Request object built with ...body spread	Explicit field assignment
A03	XSS (Taint)	dangerouslySetInnerHTML={{ __html }} without sanitizer	DOMPurify.sanitize() at boundary
A03	XSS (Taint)	Function arg → innerHTML without sanitization	Sanitize at function boundary
A03	XSS (Taint)	<a href={variable}> with user-controlled URL	Protocol allowlist (https/http only)
A01	Open Redirect (Taint)	URL param → location.href without validation	Domain allowlist or relative-only
A04	Insecure Design	postMessage handler without origin check	Strict event.origin comparison
A02	Sensitive Data Exposure	JWT stored in localStorage/sessionStorage	httpOnly cookie instead
LLM01	Prompt Injection (LLM)	Untrusted or caller-controlled value (RAG docs, fetched content, tool results, role args) reaches the prompt with no data-only framing	Delimit untrusted content as data; map caller values to fixed enumerated instructions

Reporting

The severity scale is critical / high / medium. Always include file:line. Report each independent vulnerability as its own finding. When one file holds two distinct issues (e.g. a path traversal and a separate prompt injection), list both as separate findings rather than folding one into a note on the other.

Signal	Severity	Required output
Certain exploit	critical	Full exploit scenario + concrete fix
Clear vulnerability	high	Attack vector + concrete fix
Possible issue	medium	verification_hint + suggested fix
Speculative only	none	Do NOT report

References

Topic	Scope	File
Basic	A01, A02, A07	${CLAUDE_SKILL_DIR}/references/owasp-basic.md
Injection	A03	${CLAUDE_SKILL_DIR}/references/owasp-injection.md
Advanced	A04-A06, A08-A10	${CLAUDE_SKILL_DIR}/references/owasp-advanced.md
Cloud	IAM, IaC, CI/CD	${CLAUDE_SKILL_DIR}/references/cloud-infrastructure.md
Frontend	Taint analysis	${CLAUDE_SKILL_DIR}/references/frontend-taint-checklist.md