Security Audit

/security-audit — OWASP-Coverage Security Audit

Comprehensive security scan with OWASP Top 10 (Web 2021) + OWASP GenAI/LLM Top 10 (2025) coverage. Diagnostic + automated fixes for safe issues + remediation plan for the rest. No effort estimates — security agent does not size remediation work; user/PM owns sizing per Q2.

When to use

• Pre-release security audit
• Compliance check (SOC 2, ISO 27001, PCI-DSS)
• Threat modeling exercise
• Vulnerability scan after dependency update
• Periodic (quarterly) security health check

Not for

• Code-review security feedback (line-level) → /code-review (uses code-review skill with security-checklist companion)
• Penetration testing → professional pen-test service
• Production incident response → on-call SRE + /analyze-prod

Invocation

/security-audit
/security-audit --scope "secrets,deps,auth" --report-type detailed
/security-audit --scope all --report-type summary

Arguments

Output

• <repo>/.ai-skills-memory/security-audits/<run-id>/SECURITY-REPORT.md — findings by category with severity:
  • Secrets scan (hardcoded creds, API keys) — CRITICAL if found
  • Dependency audit (CVE, outdated libs) — HIGH
  • Auth review (credentials, sessions) — MEDIUM
  • Access control (RBAC, data scoping) — MEDIUM
  • Data handling (PII exposure, encryption, logging) — MEDIUM
  • Cryptography (algorithm choices, key management) — MEDIUM
  • Infrastructure (network policies, TLS, secrets store) — MEDIUM
• <repo>/.ai-skills-memory/security-audits/<run-id>/REMEDIATION-PLAN.md — per finding: severity, mitigation steps, suggested owner role. No effort estimate (Q2)
• Pull request with fixes for automatable issues (dep updates, secret removal from code)

Agent roster

Pipeline

┌─ Scope parsing: audit --scope areas
│
├─ security-engineer + devops-engineer in parallel:
│  ├─ security-engineer:
│  │  - Grep for secrets (hardcoded keys, tokens, passwords) — see pii-patterns.txt
│  │  - Dependency scan (npm audit, pip check, cargo audit, etc.)
│  │  - Auth flow review (login, session, token, CSRF)
│  │  - Access control review (data scoping, RBAC)
│  │  - Cryptography review (algorithms, key rotation, TLS)
│  │  - Per OWASP coverage matrix below (G3)
│  │  → SECURITY-FINDINGS.md per scope
│  │
│  └─ devops-engineer:
│     - Infrastructure review (network policies, secrets store, TLS)
│     - Container scanning (if Docker)
│     - CI/CD security (secret management, artifact verification)
│     → INFRA-FINDINGS.md
│
├─ Lead consolidates → SECURITY-REPORT.md
│
├─ Lead generates REMEDIATION-PLAN.md:
│  └─ Per finding: severity, mitigation steps, suggested owner role
│     (NO effort estimate — user/PM owns sizing per Q2)
│
└─ developer (if applicable): fix automatable issues in a separate PR:
   - Update dependencies with security patches
   - Remove hardcoded secrets (replace with env-var refs + add to secret store)
   - Add missing security headers (CSP, X-Frame-Options, HSTS)
   Memory write: L4 audit summary; if CRITICAL → committed incident record

No RALF — audit is pass-once.

OWASP coverage (G3)

Eval rubric (plugin/eval/judge-rubrics/security-audit.md) MUST verify coverage of: OWASP Web Top 10 (2021) + OWASP GenAI/LLM Top 10 (2025) — see @owasp-coverage.

If the audit target includes any AI/LLM component (agent harness, prompt template, LLM API integration, RAG pipeline), the GenAI Top 10 IS in scope by default.

Methodology — SAST / DAST / SCA / IAST

security-engineer covers SAST + SCA + secret scanning. DAST runs only when an instance is reachable (typically follow-up against staging). IAST is mention-only.

Supply chain

SBOM tooling + SLSA levels + Cosign + EPSS/KEV patterns — see @supply-chain-security.

G7 spawn payloads

All spawns use structured G7 payloads. security-engineer is read-only by frontmatter (disallowedTools: Write, Edit); developer fixes happen in a separate phase with explicit user approval.

Eval rubric

Pointer: plugin/eval/judge-rubrics/security-audit.md (B10).

Dimensions:

1. Completeness — all requested scopes covered
2. OWASP coverage — Top 10 Web + GenAI Top 10 systematically addressed where applicable
3. Accuracy — findings are real issues, not false positives
4. Clarity — each finding explains the risk and impact
5. Prioritization — severity correctly assessed (CRITICAL > HIGH > MEDIUM > LOW)
6. Remediation guidance — mitigations are specific and testable

Pass: avg ≥ 4.0, no dimension < 3.

Memory writes

pre-tool-use-committed-write.py hook enforces .committed/security/incidents/*.md matches allowlist.

Failure modes

• False positive (e.g., secret string in test fixture comment): security-engineer reviews, confirms, removes from report if benign — but ALWAYS documents the false-positive review trail
• Dependency CVE has no fix available: documented in REMEDIATION-PLAN with risk acceptance note + escalation deadline
• Auth flow too complex to audit in one pass: security-engineer flags as needs-deeper-review; recommends threat-modeling session via /spike

Observability events

• workflow_start — security-audit + scope
• finding_identified (per finding + severity)
• automatable_fix_applied (per fix in developer phase)
• workflow_end — COMPLETE with critical/high/medium/low counts

Integration

• Orchestrator: feature-design-lead
• Primary agent: security-engineer (the B5 agent — read-only, OWASP-aware)
• Schemas: plugin/schemas/spawn-payload.schema.json, plugin/schemas/return-contract.schema.json
• Sub-workflow: developer fix phase uses /develop for the remediation PR
• Companions: /subagent-spawn, /context-load
• Rules: subagent-isolation, untrusted-content-wrapping (CRITICAL — audit reads project files which may have hostile content), memory-discipline (CRITICAL incidents → committed)
• Hooks: pre-tool-use-committed-write.py (enforces incident allowlist), tool-output-normalize.py (G2 on npm audit / pip check / cargo audit outputs)
• Templates: plugin/hooks/scripts/pii-patterns.txt (secret-pattern source)