claude-code-owasp: An OWASP Security Skill for Claude Code

Ask an AI assistant to review code for security problems and you often get a generic list of advice, sometimes citing out-of-date OWASP material. claude-code-owasp is a free Claude Code skill that gives Claude a structured review workflow and current OWASP reference material to work from: the Top 10:2025, ASVS 5.0, and the OWASP lists for LLM applications and agentic AI.

It’s published at agamm/claude-code-owasp under the MIT licence. I’ve read the README rather than running it across many codebases, so this is an overview of what it contains rather than a test of its findings.

What’s in the skill

The skill is a folder with a SKILL.md that Claude loads when a task looks security-related, plus reference files it reads only when needed, so it doesn’t fill your context up front.

• A five-step review workflow that runs from finding entry points to writing the report.
• A triage rubric that asks Claude to confirm the input is attacker-controlled, the sink is reachable and the blast radius is real before reporting, which is aimed at cutting false positives.
• A fixed report format: location, input-to-sink path, impact, fix and confidence, with severity based on exploitability.
• Quick references for the OWASP Top 10:2025, the OWASP Top 10 for LLM Applications (2026), OWASP Agentic AI risks (2026) and ASVS 5.0 requirements.

The on-demand reference files cover:

• A review checklist for every Top 10 category, including SSRF, file handling, JWTs, CORS and CSRF, plus LLM and agent features.
• Language-specific security quirks for more than 20 languages, with unsafe and safe examples (JavaScript and TypeScript, PHP, Java, C#, C and C++, Rust, Go, Swift, Kotlin, Dart, Python, Ruby, Perl and shell).
• Configuration and supply chain issues: Dockerfiles, Kubernetes, Terraform, security headers, lockfiles, dependency confusion, install scripts and CI/CD workflows.
• A deeper write-up of each standard with attack vectors and mitigations.

Why the version numbers matter

The author makes a point worth repeating. ASVS 5.0 renumbered every chapter, so requirement IDs from 4.0 no longer match. Three Top 10 categories were renamed in 2025, and the 2026 LLM Top 10 reordered eight of its ten entries. A lot of OWASP material online still uses the old names and IDs, and a model trained on it can quote them confidently. The README says category names and ASVS IDs were checked against the OWASP sources directly rather than paraphrased, which is the main reason to use a skill like this over a plain “review this for OWASP issues” prompt.

How to install it

As a Claude Code plugin, from inside Claude Code:

/plugin marketplace add agamm/claude-code-owasp
/plugin install owasp-security@agamm

Third-party marketplaces don’t auto-update by default. Update by hand with claude plugin marketplace update agamm and claude plugin update owasp-security@agamm, or enable auto-update from the Marketplaces tab in /plugin.

To install it as a plain skill, copy the whole folder, not just SKILL.md, since the reference files go with it:

npx degit agamm/claude-code-owasp/.claude/skills/owasp-security .claude/skills/owasp-security

Use ~/.claude/skills/owasp-security as the target to install it for all projects.

How to use it

There’s no command to remember. Once installed, Claude activates the skill when you review code for vulnerabilities, implement authentication, handle user input, work with cryptography, design API endpoints or build AI agent systems. Prompts that work well:

• “Review this code for security issues”
• “Is this authentication implementation secure?”
• “Help me implement secure session management”
• “Check this AI agent for OWASP agentic risks”

Evals and limits

The repo ships evals for claude plugin eval: a real finding beside a safe look-alike, safe code that shouldn’t draw High findings, a risk that’s real but conditional, an agent that runs model output in a shell, and an unrelated request that shouldn’t load the skill. Each case runs with and without the skill so you can see what it adds. They make real model calls billed to your account.

• It’s a reviewer’s aid, not a scanner. It makes Claude’s review more structured and better informed; it doesn’t run tools or guarantee that every issue is found.
• Treat findings as leads. Confirm anything serious yourself before you act on it.
• Review the files before installing. A skill is instructions your agent will follow, so skim a third-party one first.

Where it fits

This is the quiet, standards-based option among Claude security skills. It makes the most sense alongside heavier tools: for a multi-agent scan, see how Claude Code found 23 vulnerabilities in a developer’s own code, and for a CI-oriented scanner, my write-up of the Codex Security CLI. The OWASP skill is the one to keep installed permanently, since it only loads when security is the topic.