| 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
- Read
references/review-workflow.md first.
- Select the app-layer lane from
references/app-api-data.md.
- Add the trust-boundary lane from
references/authn-authz-session.md, references/ai-agent-mcp.md, or references/infrastructure-supply-chain.md.
- Add one language lane from
references/languages/ and one platform overlay from references/platforms/ when they apply.
- 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.mdreferences/languages/javascript-typescript.mdreferences/languages/python.mdreferences/languages/dotnet-csharp.md
Platform overlays
references/platforms/github-actions.mdreferences/platforms/aws-core.mdreferences/platforms/vercel-nextjs.mdreferences/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.
