OWASP Security Skill
Injects the latest OWASP standards (2025–2026) into code review and AI app development, so your model builds secure code instead of patching it afterward.
A pure read-only reference knowledge pack: no tool permissions, no scripts, no external effects, so least privilege is satisfied by construction; sources are explicitly attributed (owasp.org Top10:2025, OWASP/ASVS GitHub, genai.owasp.org) and the license is MIT. Deducted because SKILL.md declares no user-confirmation or rollback mechanisms and attribution exists only as links, which a static review cannot fully verify against official text.
Content is self-consistent (unsafe/safe pairs, ASVS IDs vs chapter structure agree internally) and the progressive-disclosure design is sound; no runtime dependencies exist. Deducted because static review provides no execution evidence, no tests or CI cover key paths, and some snippets (e.g. llm.complete_/QuerySpec) are illustrative pseudocode that may not run verbatim; scored at the static cap of 10.
Scenarios and triggers are clearly declared in the description (security review, authn/authz, input handling, agentic AI), with explicit non-fit boundaries ('a pattern match is not a vulnerability' triage, 'entry points, not complete coverage'), which helps accurate invocation and reduces false positives. Deducted because content is English-only with no Chinese support, and depth limits per standard are not declared. No dependence on overseas services, so mainland reachability is unaffected.
Good information architecture (always-loaded quick reference plus on-demand deep references), clear README with install, usage examples, contributing, license, and sources. Deducted for missing versioning, changelog, known-limitations list, and any declared maintenance/update path despite standards content that will age; publisher identity is unverified.
Marginal value is clear as security-review context: ASVS 5.0 renumbering guidance and the three-question false-positive filter are hard to assemble quickly by hand, and output is directly usable review guidance. Deducted because no sample outputs or validation evidence exist and the 20+ language coverage is thin (two examples each); scored at the static cap of 6.
Key claims (Top 10:2025 category names, ASVS 5.0 chapters/IDs, LLM/Agentic lists) cite specific auditable official sources with a verbatim-claim statement, and fact vs inference is well separated. Deducted because a static read cannot independently confirm the cited IDs and levels, and there is no third-party execution or cross-verification evidence; scored at the static cap of 5.
- Static review only, nothing was executed; re-verify ASVS 5.0 requirement IDs and levels against the official standard before using them in formal assessments.
- Content is English-only; the skill has no online-service dependencies and works offline.
- reference/languages.md gives only a few examples per language — the author states these are 'entry points, not complete coverage'; do not treat it as an exhaustive checklist.
- Some LLM examples (complete_/QuerySpec) are illustrative pseudocode and will not run verbatim.
- No versioning or changelog; OWASP/ASVS content may age as standards update; publisher identity is unverified.
- No signs of malicious code, credential theft, or over-broad permissions; no red-line issues found.
What does this skill do, and when should you use it?
A Claude Code skill that condenses OWASP Top 10:2025, ASVS 5.0, the LLM Top 10 (2025), and Agentic AI security (2026) into an always-loaded core instruction file, backed by on-demand reference documents. It provides vulnerability quick-reference tables, security code review checklists, unsafe/safe code patterns, and a finding-triage rubric designed to cut false positives by verifying attacker-controlled input, sink reachability, and blast radius before reporting. Category names and ASVS requirement IDs are verified directly against owasp.org and genai.owasp.org rather than paraphrased from secondhand sources. It suits developers working across both traditional web security and AI/RAG/agent systems.
The skill works as pure Markdown instructions: Claude Code activates it automatically when you review code for vulnerabilities, implement authentication/authorization, handle user input, or build AI agent systems. It supplies the OWASP Top 10:2025 quick table, review checklists for input handling/auth/sessions/access control/data protection/error handling, secure patterns for SQL and command injection prevention, Argon2 password storage, and fail-closed error handling, plus mitigation tables for LLM01–LLM10 and ASI01–ASI10 risks. Deeper content (security quirks for 20+ languages, per-item attack vectors for each standard) lives in reference/ files loaded on demand. It runs no scripts or commands; it produces review findings and fix recommendations only.
- A developer asks the model to review a code snippet for injection, access control, or cryptographic flaws before committing
- A backend engineer implementing login/session management wants the model to verify against specific ASVS 5.0 requirement IDs (e.g. password length 6.2.1, session entropy 7.2.3)
- A team building a RAG chatbot or tool-calling app checks for prompt injection (LLM01) or excessive agency (LLM06) risks
- An AI agent developer audits a system against the OWASP Agentic AI 2026 ASI01–ASI10 risk list
- A security team wants review output limited to genuinely exploitable findings, not pattern-match false positives, via the skill's three-step triage rubric
- A full-stack developer looks up common vulnerabilities and safe patterns for a specific language among 20+ covered (Python, Java, Rust, etc.)
What are this skill's strengths and limitations?
- Covers four standards in one install: OWASP Top 10:2025, ASVS 5.0, LLM Top 10 (2025), and Agentic AI (2026)
- IDs and category names verified directly against owasp.org and genai.owasp.org — important because ASVS 5.0 renumbered everything and much circulating material still cites 4.0 IDs
- Built-in triage rubric (attacker-controlled input, sink reachability, blast radius) explicitly instructs severity by exploitability, targeting the most common false-positive failure mode
- All code examples shown in unsafe/safe pairs, easy to imitate
- Progressive disclosure design: core file always loaded, 20+ language details read on demand, conserving context
- Pure Markdown with no script dependencies; MIT licensed and freely modifiable
- No automated tests or benchmark data demonstrating review accuracy; effectiveness depends entirely on the underlying model
- The author makes no claim of validation against current model versions, and compatibility with other editors/tools is unverified
- It is knowledge injection only — no substitute for real SAST/DAST tooling or manual penetration testing
- ASVS Level 3's 92 requirements are not enumerated; a formal L3 assessment still requires the standard itself
- The depth of the reference files (languages.md, owasp-report.md) cannot be verified from SKILL.md alone
How do you install this skill?
Copy the skill directory into your project with degit:
npx degit agamm/claude-code-owasp/.claude/skills/owasp-security .claude/skills/owasp-security
Or install globally for all projects:
npx degit agamm/claude-code-owasp/.claude/skills/owasp-security ~/.claude/skills/owasp-security
Alternatively, git clone the full repository and copy the .claude/skills/ folder manually. You must copy the entire folder (SKILL.md plus the reference/ subdirectory), since the skill relies on progressively disclosed reference files.
How do you use this skill?
Once installed, trigger it with natural prompts in Claude Code, e.g. "Review this code for security issues", "Is this authentication implementation secure?", or "Check this AI agent for OWASP agentic risks". It activates automatically in security review, auth, input handling, cryptography, and API design contexts. Deep content loads on demand: the model reads reference/languages.md when reviewing a specific language, and reference/owasp-report.md for per-item attack vectors of any covered standard.