メニュー
Runs a full OWASP/CWE security audit via isolated subagents. Use when the user types /security-review or explicitly requests a full security audit of the current code.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Detects supply-chain risks in dependency manifests, lockfiles, install commands, and CI pipelines — missing lockfiles, absent vulnerability scanners, unverified installs, install hooks that pipe remote content to a shell, and AI-hallucinated package names. Use when writing package install commands, dependency manifests, CI/CD configs, build scripts, or code that loads external packages.
Deep review that audits API contracts for mismatches between what callers assume and what implementations enforce. Designed for nightly CI, pre-release scans, or manual deep audits — runs in tens of minutes, not seconds. Surfaces bugs that single-pass OWASP review misses — caller/callee invariant gaps, trust-anchor confusion, predicate misnaming.
Detects weak password storage, flawed JWT validation, sessions surviving logout, and missing MFA in authentication code. Use when writing login flows, session management, password storage, JWT handling, MFA implementation, credential validation, or any code that authenticates users. Also invoke when implementing "remember me" functionality or API key management.
Detects missing ownership checks, broken role enforcement, and IDOR vulnerabilities that let users access other users' resources. Use when writing code that checks authorization, enforces resource ownership, handles IDOR (object-level authorization), processes server-side requests to external URLs, or implements access control middleware.
Detects multi-threaded code where a lock is held across a blocking operation, lock acquisition order risks deadlock, lock-free atomics use the wrong memory ordering, or double-checked locking is missing the necessary barrier. Use when writing or modifying code that acquires multiple locks, uses atomic operations with explicit memory orders (std::memory_order_*, atomic.LoadAcquire, Ordering::Relaxed), implements lazy initialization with double-checked locking, or calls a blocking operation (I/O, sleep, await, blocking channel send) inside a locked region.
Detects library-internal cryptographic correctness bugs that pattern matchers and crypto-failures skills miss — AEAD nonce reuse, ECDSA k-value reuse, length-extension on bare hashes, padding-oracle exception distinguishability, and branching on secret material. Use when writing code that calls a crypto primitive directly (AEAD encrypt/decrypt, ECDSA/DSA sign, raw hash, RSA decrypt). Distinct from cryptographic-failures, which covers application-layer choices like MD5 for passwords; this skill is about how primitives are wired up.
| name | security-review |
| description | Runs a full OWASP/CWE security audit via isolated subagents. Use when the user types /security-review or explicitly requests a full security audit of the current code. |
Full repo audit against OWASP Web + LLM Top 10:2025 via a four-stage
pipeline: threat-model → hotspots → review → validate. Main context
dispatches subagents (threat-modeling, hotspot-mapping,
vulnerability-audit, design-review, finding-validate,
attack-chain-analysis) and renders findings; never reads code.
This skill is the orchestrator. The actual pattern catalog lives in
the per-category auto-invoking skills (injection, csrf, ssrf,
broken-access-control, authentication-failures, etc.) — the
vulnerability-audit subagent picks the right one per hotspot and
applies its Vulnerable patterns section. Skill-list maintenance
is automatic via .claude/skills/ directory contents; no separate
catalog file.
Use only the Agent tool in main context. No Read/Grep/Glob/
Bash in main context. Stage prompts live in .claude/agents/:
threat-modeling, hotspot-mapping, design-review,
vulnerability-audit, finding-validate, attack-chain-analysis.
This skill is just the coordinator.
Copy this checklist as you progress:
- [ ] Stage 0 — threat-modeling returned
- [ ] Stage 1 — hotspot-mapping returned (one whole-repo call)
- [ ] Stages 1b+2 — design-review + N vulnerability-audit in ONE message
- [ ] Stage 2.5 — finding-validate returned; refuted findings dropped
- [ ] Stage 3 — attack-chain-analysis returned
- [ ] Stage 4 — findings table rendered with severity legend
- [ ] Stage 5 — suggested /security-cleanup to the user
Dispatch one threat-modeling subagent. It returns JSON with
purpose, deployment, trusted_inputs, untrusted_inputs. Thread this
into every later subagent.
Dispatch one hotspot-mapping subagent with the threat model.
Returns a JSON array of {file, lines, name, category, priority, why}
entries — the hotspot list for Stage 2.
Collapse hotspots by file first — group into {file: [hotspots]}.
Then in a SINGLE message, dispatch:
design-review subagent with the threat model.vulnerability-audit subagent per file with the threat
model AND that file's hotspot list. trusted_inputs /
untrusted_inputs informs which sinks are reachable. Serial
launches defeat parallelism.Concatenate every returned findings array. Dedupe by (file, line).
Dispatch one finding-validate subagent with the merged findings
array and the threat model JSON. It returns a JSON array of 0-based
indices to drop — findings refuted by concrete evidence at the
cited line (guard, sanitizer, correct API). [] is common. Remove
the indicated indices before Stage 3.
Dispatch one attack-chain-analysis subagent with the threat model
AND the merged findings. The threat model informs effective severity
(does the chain cross an untrusted boundary?). Returns chain objects
with plain-English narratives, or [].
Emit # Security Review, then the legend Critical = anyone on the
internet can exploit. High = needs an account. Medium = limited blast
radius. Low = defense-in-depth. Then a findings table:
Severity / Where / What's wrong / How to fix (use the auditor's
finding/fix verbatim; append (category: <skill>) to What's
wrong). If chains exist, emit ## Attack chains with
### Chain N — <effective_severity> per chain as prose. One summary
line. Zero findings: Security review complete. No findings across N hotspots.
Suggest /security-cleanup to apply fixes. Do not auto-rewrite.
severity, file:line, OWASP/CWE category, and a concrete fixfinding/fix text is plain language a non-security developer can act on