audit-security
This skill contains shell command directives (!`command`) that may execute system commands. Review carefully before installing.
!cat ~/.claude/skills/audit-workflow.md
Run as the security dimension. Lens:
Find vulnerabilities where untrusted input or weak controls let an attacker change behavior, exfiltrate data, or escalate privilege. Tag each finding with a CWE.
Under-weighted without prompting: shell or subprocess called with string args instead of array (command injection); SQL via string concatenation, template literals, or ORM escape hatches (.raw(), .extra(), RawSQL) — parameterized queries non-negotiable; deserialization of untrusted input (pickle, yaml.load, JSON-into-class hydrators) and XML parsers at default settings (XXE); template engines or DOM APIs rendering user input unescaped (dangerouslySetInnerHTML, innerHTML, unescaped template vars); broken access control — routes w/o authentication check, state-changing endpoints missing CSRF protection, IDOR where a resource is fetched by user-supplied ID w/o ownership check; JWT misuse — alg:none accepted, key/algorithm confusion, missing exp/iss/aud validation; weak primitives — md5/sha1 for passwords, Math.random for tokens, ECB mode, missing salt; secrets in source, logs, or error responses; trust-boundary gaps — server relying on client-side validation, parsers invoked on raw input w/o size/depth limits; SSRF / path traversal where user input becomes a URL or filesystem path w/o allowlist; LLM output trusted as code, SQL, or shell (insecure output handling).
Filter aggressively. Not a vulnerability: server-controlled config (env vars, constants, settings.*), framework-mitigated sinks (React {x}, parameterized ORM, prepared statements), UUID identifiers, or client-side-only validation when a server check also exists. A finding needs a concrete attack path from an untrusted source to a sink — theoretical "could be vulnerable if" is noise.
Auto-fix when the safe pattern is already established in the codebase or by the framework: parameterize the query, switch subprocess to array form, escape the rendered output, add the ownership check, swap md5 for argon2, scrub the secret from logs — each behavior-preserving for legitimate input. Sign-off: auth/session/crypto changes that rotate secrets, migrate algorithms, or change token format — they affect active sessions and stored credentials; sketch the change and surface with the CWE + attack path closed.