security-review

name

security-review

description

Use this skill whenever code changes touch authentication, authorization, sessions, secrets, untrusted input, API endpoints, AI/agent integrations, CI/CD, cloud infrastructure, dependency supply chain, or other sensitive trust boundaries. It routes the agent to deeper review lanes by workflow, app layer, language, platform, and zero-day / assume-breach posture while preserving the existing benchmark tooling.

Security Review

Use this skill to turn a vague “check security” request into a focused review. Do not try to carry the whole checklist in this file; route immediately to the narrowest deep references that match the change.

Start here

1. Read references/review-workflow.md first.
2. Select the app-layer lane from references/app-api-data.md.
3. Add the trust-boundary lane from references/authn-authz-session.md, references/ai-agent-mcp.md, or references/infrastructure-supply-chain.md.
4. Add one language lane from references/languages/ and one platform overlay from references/platforms/ when they apply.
5. If the change expands blast radius or depends on partial trust, also read references/zero-day-resilience.md.

Lane selector

Core workflow

• references/review-workflow.md — default review sequence, evidence expectations, and exit criteria.
• references/zero-day-resilience.md — assume-breach, zero-trust, resilience, rollback, and containment guidance.

Identity and session boundaries

• references/authn-authz-session.md — authentication, authorization, session lifecycle, tokens, cookies, impersonation, tenancy, MFA.

Application layers

• references/app-api-data.md — input validation, files, APIs, SSRF, deserialization, database access, queues, caches, logging, privacy.
• references/ai-agent-mcp.md — prompt injection, tool misuse, capability scoping, MCP safety, retrieval poisoning, output handling.
• references/infrastructure-supply-chain.md — CI/CD, artifact trust, SBOMs, provenance, secrets delivery, cloud controls, incident hooks.

Language lanes

• references/languages/go.md
• references/languages/javascript-typescript.md
• references/languages/python.md
• references/languages/dotnet-csharp.md

Platform overlays

• references/platforms/github-actions.md
• references/platforms/aws-core.md
• references/platforms/vercel-nextjs.md
• references/platforms/supabase.md

Worked examples

• assets/examples/pr-code-review.md — worked pull-request review with evidence and review comments.
• assets/examples/threat-model-assume-breach.md — compact threat model using assume-breach posture.

Automated evidence layer (review-cli security)

This skill routes and judges; it does not hand-run scanners. When the code-review family is published to this home, feed it deterministic evidence from the review-cli security command family (Spec #107 local-devsecops-hardening) and fold the findings into the lanes above:

• review-cli security audit --target <dir> — DevSecOps posture gap inventory (SAST, secret, dependency-vuln, SBOM, signing, OWASP-LLM agent-safety) → references/infrastructure-supply-chain.md.
• review-cli security scan --target <dir> [--delegate] — multi-language SAST + secret + dependency-vuln + SBOM; --delegate routes to the aclab-middlewares security-stack (sectool → Dependency-Track/DefectDojo) when reachable, else local tools → the language + supply-chain lanes.
• review-cli security grounding --transcript <f> — OWASP-LLM content-safety + hallucination gate (fabricated refs LLM09, secret leakage LLM06, prompt-injection LLM01) → references/ai-agent-mcp.md. Install as a Stop-hook on any agent: review-cli security grounding --emit-hook <claude|codex|kiro|opencode> prints the ready-to-install config (Claude/Codex = JSON Stop hook; Kiro = hooks.stop; opencode = session.idle JS plugin). To install it automatically (safe merge + backup + idempotent) run python3 scripts/install_grounding_hooks.py --agent all. The installer defaults to each agent's flat project skill home, such as ~/.claude/skills/code-review or ~/.codex/skills/code-review; use --binary, --skill-dir, GROUNDING_CODE_REVIEW_SKILL_DIR, or REVIEW_CLI_SKILL_DIR when a target home uses a different layout or a shared binary.

These are evidence, not verdict — this skill (and the change's own review) keep authority. Absent tools degrade honestly (tool-unavailable / local-fallback); record that, never silent-skip.

Complementary persona: gstack CSO

For periodic, org-wide, infrastructure-first security-officer audits (secrets archaeology, supply chain, CI/CD, skill supply-chain scanning, OWASP + STRIDE, daily/comprehensive modes, trend tracking), compose with the gstack cso skill (https://github.com/garrytan/gstack) when installed. Division of labour: this security-review skill is the per-change trust-boundary reviewer; cso is the periodic whole-system audit persona. Both consume the same review-cli security evidence above. Route the monthly deep pass to cso; do not duplicate its workflow here.

Maintenance / provenance

This skill is repo-owned by aclab-code-review-private (.agents/specs/local-devsecops-hardening, Spec #107) and published to the global skill homes as a code-review companion via scripts/publish_code_review_skill.py. Edit it here (the source of truth), not in a global home.

Minimum output contract

Return a review that is specific to the changed trust boundaries, not a generic wall of bullets.

• State what was reviewed.
• Name the main risks that are real for this change.
• Separate confirmed issues from “verify before merge” concerns.
• Give concrete remediations with file or layer hints.
• Note what evidence was checked: tests, policies, config, code paths, or deployment settings.

Benchmarking and provenance

Keep the existing benchmark tooling intact.

• Benchmark runner: scripts/run_security_benchmark.py
• Local benchmark docs: evals/BENCHMARKING.md
• Copy alignment helper: scripts/sync_copies.py

Do not write benchmark state into the installed skill directory. Preserve the existing provenance behavior instead of inventing a new storage policy.

Compatibility note

cloud-infrastructure-security.md is retained as a stub for older references. Treat the deep infrastructure guidance in references/infrastructure-supply-chain.md plus the platform overlays as the canonical source.