security

Invoke the security skill on the diff and it maps actors, trust boundaries, and data flows before checking authorization placement, input validation, secrets handling, and dependency risk. Findings that enable unauthorized access or data exposure block merge unless explicit risk acceptance is recorded.

It verifies PKCE with S256, per-request state and nonce validation, exact redirect_uri matching, refresh-token rotation, and audience binding. For token validation it checks JWKS pinning, algorithm allowlists rejecting alg:none, and iss/aud/exp/nbf claims.

Yes, the ai-agent reference treats all external content channels as untrusted input and every tool call as a privileged action. It covers capability isolation, schema-validated tool parameters, output sanitization against markdown exfiltration, and RAG index poisoning defenses.

It selects the auditor by lockfile: pnpm audit, npm audit, pip-audit, cargo audit, govulncheck, bundler-audit, composer audit, and others per ecosystem. For multi-language repos it falls back to osv-scanner, and recommends trivy or grype for container image scanning.

Skip it for general code quality work with no trust boundary, API shape design without security semantics, and runtime alert design, which belong to other skills. It also assumes networked applications; embedded, firmware, and mobile binaries need platform-specific guidance.

It requires rotating the credential first, revoking derived sessions, auditing usage logs, then scrubbing history with git filter-repo. Detection uses gitleaks or trufflehog in CI with redacted output, and blocking merge on any new finding.