Skip to main content
Guide8 min read·Updated July 27, 2026
🧩

Best AI Agent Skills for Penetration Testing in 2026

B

A. Frans

Published July 27, 2026

Agent SkillsPenetration TestingSecurityClaude CodeRed Team

I went through every offensive-security entry in our skills directory and checked what each one ships. Of the seven most-starred results, two are real Claude Code Agent Skills. The rest are an MCP server, a standalone CLI, a deliberately broken training target, and a markdown reference repo.

That matters because the install command is different for each, and copying the wrong one gets you a command-not-found error at best. At worst you pipe a shell script from a URL into bash on a machine that holds client engagement data.

So this list is organized by what each project ships, verified against each project's own repository on 27 July 2026.

What's real, at a glance

ProjectIs it a Claude Code skill?StarsLast pushedUse it for
Trail of Bits SecurityYes, a plugin marketplace of 50+ skills6,2742026-07-20Code audit, smart contracts, variant analysis
ProwlerYes, ships a skills/ directory14,4842026-07-24Cloud posture, compliance scanning
ffuf ScannerYes, a single manual-install skill1982025-10-16Web fuzzing, content discovery
HexStrike AINo, it's an MCP server10,5042026-04-27Orchestrating 150+ existing security tools
StrixNo, it's a standalone CLI44,4772026-07-26Autonomous app pentesting in CI
OSINT Cheat SheetNo, it's a reference repo2,0952026-07-22Recon methodology, tool lookup
Damn Vulnerable MCP ServerNo, it's a target to attack1,3212025-12-08Practicing MCP exploitation
Star counts and push dates come from the GitHub API on the date above. The "is it a skill" column comes from reading each repository's own README and file tree, not from a directory listing, including ours.

The two you should install first

Trail of Bits Security

This is the strongest offensive-security package available for Claude Code right now, and it isn't close. Trail of Bits is a security firm with a long public audit record, and they've packaged their internal review methodology as a plugin marketplace.

Install it the documented way:

/plugin marketplace add trailofbits/skills

Then browse and install individual plugins from that marketplace. The catalog runs past 50 entries across smart contract security, code auditing, malware analysis, verification, reverse engineering and mobile security. Named examples in their README include building-secure-contracts, rust-review, c-review, semgrep-rule-creator and variant-analysis.

Variant analysis is the one worth calling out. You give it a known bug and it hunts for the same class of mistake elsewhere in the codebase. That's the highest-value pattern in a real audit and the one most likely to be skipped under time pressure.

The license is CC-BY-SA-4.0, which is unusual for a code repository and worth reading if you plan to redistribute anything derived from it.

Prowler

Prowler is a cloud security platform first and a skills provider second, but the skills side is legitimate. The repository has a skills directory and a setup script that wires it into your agent:

./skills/setup.sh

The script creates symlinks into the directories used by tools that follow the agentskills.io standard, including .claude/skills/. Read the script before running it, which is advice that applies to every setup script in this article and doubly to security tooling.

What you get is Prowler's own coding patterns, testing conventions and architecture guidance. That's useful if you're extending Prowler with custom checks. It is not a general-purpose offensive skill, and the marketing around "AI security skills" blurs that line more than it should.

For the actual scanning work, Prowler covers AWS, Azure, GCP and Kubernetes posture against a long list of compliance frameworks. Apache 2.0, actively maintained, 14,484 stars.

The single-purpose skill that still works

ffuf Scanner is 198 stars and hasn't been touched since October 2025. It's also the cleanest example of what a focused Agent Skill looks like.

Install is manual, exactly as the README documents:

  • git clone https://github.com/jthack/ffuf_claude_skill
  • mkdir -p ~/.claude/skills
  • cp -r ffuf_claude_skill/ffuf-skill ~/.claude/skills/

After that Claude can drive ffuf for directory brute-forcing, subdomain discovery and API endpoint enumeration, and interpret the output instead of dumping 4,000 lines of 403s at you. You need ffuf itself installed separately.

Nine months without a commit would worry me for a framework. For a wrapper around a stable CLI, it's fine. The skill's job is translating intent into flags, and ffuf's flags haven't changed.

The three that aren't skills, and what to do with them

Strix is a CLI, not a skill

Strix has 44,477 stars and shows up at the top of every "AI pentest skills" list, including search results that will bring you to this page. It is not a Claude Code skill. Its README documents no SKILL.md and no Claude Code integration. It's a Python CLI that runs autonomous agents against a target:

strix --target ./app-directory

The documented install is a piped shell script from strix.ai, which I'd rather you fetch, read, and run in that order. It needs Docker and an LLM API key from OpenAI, Anthropic or Google.

What it does is real and useful: dynamic testing that runs the code, validates findings with working proof-of-concept exploits rather than reporting theoretical issues, and generates patches. It plugs into CI and GitHub Actions. Apache 2.0, pushed the day before this article went up.

Use it as a pipeline stage. Don't expect a slash command.

HexStrike AI is an MCP server

HexStrike wires an AI agent to 150+ existing security tools through the Model Context Protocol. That's an orchestration layer, not a skill, and it installs through MCP configuration rather than a skills directory. Our MCP servers versus agent skills explainer covers why the distinction changes what you can do with it.

Two cautions. It was last pushed in April 2026, which is a long gap for a project whose value is keeping 150 tool integrations current. And any MCP server that can invoke arbitrary security tooling deserves a hard look at its permission surface before it touches a real engagement.

The other two are reference material

OSINT Cheat Sheet is a curated markdown collection of recon tools, datasets and methodology. Excellent as reading. Point Claude at the raw file if you want it in context. It is not installable as a skill and doesn't claim to be.

Damn Vulnerable MCP Server is a deliberately insecure MCP server built for practicing attacks against MCP itself. If you're assessing agent infrastructure rather than web apps, it's the best practice target available. Run it isolated. That is the entire point of it.

Security rules for security skills

The irony of installing unvetted code to do security work is not lost on anyone, but it keeps happening.

Read the SKILL.md before you install it. It's a markdown file with instructions that go straight into your model's context. That's the whole attack surface, and it's readable in two minutes.

Check the allowed-tools frontmatter. A skill can pre-approve tools for the turn that invokes it, which means it can grant itself the ability to run bash without prompting you. Claude Code's own documentation warns that project skills need the workspace trust dialog accepted first, and tells you to review them before trusting a repository. Follow it to the letter on offensive tooling.

Never install from a directory's install command without checking the source repo. Including ours. Directory metadata goes stale, and install syntax in this ecosystem has changed more than once. The repository README is the primary source. We wrote up the full process in how to audit agent skill security, and our sandbox-or-install decision tree covers when to isolate rather than trust.

Keep engagement data off the machine running experimental skills. Client scope documents, credentials and findings do not belong on the same box where you're trying out a 200-star GitHub project.

Test on authorized targets only. Damn Vulnerable MCP Server, your own lab, or systems covered by a signed scope. An AI agent that can chain tools autonomously makes it much easier to accidentally touch something outside scope, and "the agent did it" has never once worked as a defense.

What's still missing

The gaps in this category are large enough to be worth naming.

There is no strong Agent Skill for report writing, which is where pentesters lose their weekends. There's nothing good for scope management, so nothing stops an agent from wandering past a boundary. Mobile testing is thin outside the Trail of Bits catalog. And the OSINT side is entirely reference material with no executable skill behind it.

If you're looking for a skill to build and publish, any of those four is uncontested. Our guide to building your own skill walks through the structure, and the broader cybersecurity skills roundup covers the defensive side. For tooling outside the agent ecosystem, see our full list for cybersecurity professionals.

FAQ

What's the difference between an Agent Skill and an MCP server for security work? A skill is a markdown file of instructions that loads into context when relevant. An MCP server is a running process that exposes tools the model can call. Skills teach methodology. MCP servers provide capability. Trail of Bits is the former, HexStrike is the latter, and you'll usually want both.

Is "claude skill add" a real install command? No. It appears in a lot of directory listings, including older entries in ours, but it doesn't exist in Claude Code's documentation. The documented paths are the plugin marketplace commands, or creating the directory yourself at ~/.claude/skills/name/SKILL.md for personal skills and .claude/skills/name/SKILL.md for project skills.

Can these skills run scans without my approval? They can if you let them. A skill's allowed-tools frontmatter pre-approves tools for the turn it's invoked in, and that grant clears when you send your next message. Read the frontmatter before installing anything that can reach a network.

Which one should I install if I only install one? Trail of Bits. Broadest coverage, strongest provenance, documented install path, actively maintained. Add Prowler if your work is cloud-side.

Do any of these replace a human pentester? No, and the useful ones don't claim to. Strix comes closest for web application testing in CI, and it's good at finding the classes of bug it knows about. It won't find the business logic flaw that requires understanding what the application is for.

Share this article

📬

Get More AI Tool Guides

New comparisons and guides every week. Join thousands of professionals staying ahead of the AI curve.