Security · Blog Post

Trail of Bits Skills: a security audit toolkit for Claude Code and Codex

Trail of Bits is a security research and auditing firm, and this repo is how it packages its audit methods for coding agents: 47 plugins covering CodeQL and Semgrep scans, variant analysis, diff review, fuzzing, smart contracts, and auditing the GitHub Actions that run AI agents. It is the most security-specific skill marketplace I have reviewed, and it assumes you know what to do with the findings.

By Mehadi Hasan 8 min read

What it is

trailofbits/skills describes itself as "Trail of Bits skills for security research, vulnerability detection, and audit workflows." It is a plugin marketplace, not a single skill. You add the marketplace once, then install the plugins you need. Each plugin bundles one or more skills, and some add subagents, slash commands, hooks, or workflow scripts.

The README says it installs in Claude Code and Codex, and that a ChatGPT workspace marketplace can import it from the repo's .claude-plugin/marketplace.json. As of November 17, 2026 the repo has about 8,250 GitHub stars and 714 forks, and its last push was on November 12, 2026. It is licensed under Creative Commons Attribution-ShareAlike 4.0, not a typical software license. More on that under pricing.

What's inside

47 Plugins in the marketplace
99 SKILL.md files
9 README categories
12 Workflow scripts

You will see very different skill counts quoted for this repo. That is partly because it grows fast and partly because people mix up plugins and skills. I counted from the repo itself on November 17, 2026: 47 plugin folders (matching the 47 entries in marketplace.json), and 99 SKILL.md files inside them. Most plugins hold one skill. A few hold many: trailmark and testing-handbook-skills have 15 each, building-secure-contracts has 11.

The README groups the plugins into nine categories: Smart Contract Security (2), Code Auditing (19), Malware Analysis (1), Verification (9), Reverse Engineering (1), Mobile Security (1), Development (12), Team Management (1), and Tooling (1). Code Auditing is the core. The plugins worth knowing first:

  • static-analysis: Three skills: codeql, semgrep, and sarif-parsing. CodeQL covers Python, JavaScript/TypeScript, Go, Java/Kotlin, C/C++, C#, Ruby, and Swift. Semgrep presents its ruleset plan for approval before it runs. SARIF parsing reads results without running anything.
  • audit-context-building: Reads unfamiliar code one function at a time and writes down what each function assumes, guarantees, and depends on. It deliberately reports no vulnerabilities; it is the groundwork for the pass that does.
  • variant-analysis: Starts from one bug you already found and hunts for the other instances of the same root cause, including turning that one instance into a CodeQL or Semgrep query.
  • agentic-actions-auditor: Audits GitHub Actions workflows that run Claude Code Action, Gemini CLI, OpenAI Codex, or GitHub AI Inference, checking nine attack vectors (A to I) where attacker-controlled input can reach the agent.
  • differential-review and fp-check: Security review of a PR or diff using git blame, caller counts, and test coverage, plus a separate skill that returns a TRUE POSITIVE or FALSE POSITIVE verdict on a suspected bug.
  • testing-handbook-skills: 15 skills drawn from Trail of Bits' Testing Handbook (appsec.guide): AFL++, libFuzzer, LibAFL, cargo-fuzz, Atheris, OSS-Fuzz, AddressSanitizer, coverage analysis, Wycheproof, and more.
  • building-secure-contracts: 11 skills: vulnerability scanners for 6 blockchains (Algorand, Cairo, Cosmos, Solana, Substrate, TON) and 5 development-guideline assistants.

Beyond those you will find a supply-chain auditor for npm, PyPI, and Go dependencies, C/C++ and Rust security reviews, a constant-time analyzer for timing side-channels in crypto code, a zeroization auditor, YARA rule authoring, a Firebase APK scanner, CVSS v4.0 scoring, and a Semgrep rule creator. The Development group is less about security: modern Python and C++ conventions, devcontainer setup, and git cleanup. There is also a plugin that draws Tarot cards "to add entropy to vague planning," so not everything is serious.

How it works

Each skill is a standard SKILL.md with a name, a description of when to use it, and an allowed-tools list. The descriptions are written to trigger on plain requests. The codeql skill fires on phrases like "run codeql" or "find vulnerabilities in this repo." The variant-analysis skill fires on "are there others like this?" right after you find a bug.

What stands out is how much the skills limit their own scope. The sarif-parsing description says it "does NOT run scans." audit-context-building builds understanding and reports no bugs. fp-check is "not for hunting or discovering new bugs." And the tool lists differ by job: sharp-edges only gets Read, Grep, and Glob, while the CodeQL skill gets Bash because it has to build a database.

The heavier plugins move the deterministic work into scripts and leave judgment to the model. The supply-chain auditor is explicit about why: its SKILL.md says hand-collected figures "were measured wrong," so collectors measure maintainers, downloads, and advisories, and the agent interprets them. Twelve plugins also ship workflow scripts that run as slash commands, such as the Semgrep scan below.

Most of these skills drive real tools you install yourself. The codeql skill checks for the codeql CLI, jq, and uv before it starts, and stops if any are missing. The Semgrep skill uses Semgrep Pro for cross-file taint analysis when it is available. The fuzzing skills assume the fuzzers.

Installation

In Claude Code, add the marketplace and browse it:

Claude Code: add the marketplace

/plugin marketplace add trailofbits/skills
/plugin menu

The marketplace is named trailofbits, so you can also install a plugin directly by name, the form the plugin READMEs use:

Claude Code: install one plugin

/plugin install static-analysis@trailofbits

Codex supports Claude plugin marketplaces directly:

Codex

codex plugin marketplace add trailofbits/skills
codex plugin list
codex plugin add <plugin-name>@trailofbits

Install only the plugins you will use. Every installed skill adds its description to the agent's context at session start, and a Tarot plugin has no business sitting next to your CodeQL setup on a client repo.

Using it day to day

Most skills trigger from a plain request. Ask for a Semgrep scan and the semgrep skill detects your languages, picks rulesets, and shows you the plan before running. If you have already decided, the workflow version skips the approval step. The README says that is safe because the scan is read-only over the target:

Slash commands from the plugin READMEs

/static-analysis:semgrep-scan {"target": "/abs/path", "mode": "run-all"}
/insecure-defaults:audit src/

The skills chain in the order an audit runs. Start with audit-context-building on unfamiliar code. Run static analysis. Push suspicious findings through fp-check. When something is real, use variant-analysis to find its siblings. For pull requests, differential-review covers the change and its blast radius. That last one overlaps with Anthropic's Claude Code Security Review. That tool is a PR gate. These skills are an analyst's bench.

For teams running AI agents in CI, agentic-actions-auditor is the plugin to start with. It flags patterns like ${{ github.event.* }} interpolated straight into a prompt, pull_request_target combined with checking out the PR's code, AI output passed to eval, and wildcard user allowlists. Those are the risks covered in my AI coding agent guardrails guide, turned into a check you can run.

Trail of Bits publishes two companion repos. skills-curated holds 29 third-party plugins that its staff reviewed before allowing them internally. The README says "published skills have been found with backdoors and malicious hooks." claude-code-config holds the firm's Claude Code defaults. It runs in bypass-permissions mode, so it relies on /sandbox plus deny rules for SSH keys, cloud credentials, and package tokens. It also points to devcontainers or disposable droplets for full isolation.

Pricing

Free. The repo is licensed CC BY-SA 4.0, and skills-curated uses the same license. You can use, adapt, and share the skills. If you redistribute a modified version, you must credit Trail of Bits and share it under the same license. That matters if you plan to fork a skill into a commercial internal pack. Your real costs are model usage, which is heavy for parallel subagent workflows on big codebases, and any paid tooling you choose to use, such as Semgrep Pro.

My take

This is the most credible security skill pack on this list, and the reason is in the text. The skills spend as much effort on what not to do as on what to do. They separate understanding from bug hunting, and bug hunting from verification. They keep tool access narrow, and they push measurement into scripts. That is how an audit firm works, and it shows. The trophy case already lists a real result: a timing side-channel in RustCrypto's ML-DSA signing, found with constant-time-analysis.

It is not a beginner tool. Most of the value assumes you can run CodeQL or a fuzzer, read SARIF, and judge whether a finding matters. For a team that ships web apps and wants a safety net, a PR-level reviewer is the easier first step. For security engineers, auditors, and anyone reviewing smart contracts or C/Rust code, it is worth installing plugin by plugin. If you run agents in CI, install agentic-actions-auditor first.

Frequently asked questions

How many skills does trailofbits/skills actually have?

Counted from the repo on November 17, 2026: 47 plugins (the same number listed in its marketplace.json) containing 99 SKILL.md files. Most plugins hold a single skill; trailmark and testing-handbook-skills hold 15 each and building-secure-contracts holds 11. Counts quoted elsewhere differ because the repo keeps growing and because plugins and skills get mixed up.

Do I need CodeQL or Semgrep installed?

For the static-analysis plugin, yes. The skills drive the real tools rather than imitating them: the codeql skill checks for the codeql CLI, jq, and uv before it starts and stops if one is missing, and the Semgrep skill runs semgrep and uses Semgrep Pro for cross-file taint analysis when available. Plugins like audit-context-building, sharp-edges, or fp-check only need the agent itself.

Does it work outside Claude Code?

The README documents Claude Code (/plugin marketplace add trailofbits/skills) and Codex (codex plugin marketplace add trailofbits/skills, then codex plugin add <name>@trailofbits). It also says a ChatGPT workspace marketplace can import the repo's .claude-plugin/marketplace.json.

What are skills-curated and claude-code-config?

Two companion repos from Trail of Bits. skills-curated is a marketplace of third-party plugins its staff reviewed before allowing internal use, because published skills have been found with backdoors and malicious hooks. claude-code-config is the firm's Claude Code setup: sandboxing, deny rules for credentials, hooks, and recommended skills and MCP servers.