Security Audit skill for Claude Code: audit a codebase for vulnerabilities

Cloudflare's skill that makes Claude run a structured security audit of a codebase and report verified vulnerabilities.

Installs: 73.4K; on skills.sh
Stars: 56.2K; on GitHub
Files: 19; besides SKILL.md
Updated: Sep 2026; last repo commit

Ask Claude to "security audit this codebase" and it runs a six-phase audit: reconnaissance, coverage-led hunting, candidate validation, structured output, independent verification and a report. A fresh agent tries to disprove every finding before it counts. For smaller asks, such as a focused review or a security question, it stays in guidance mode and writes no files.

A full audit writes its results outside your repo, by default in ~/security-audit-skill/<repo-name>/run-<N>, as REPORT.md, FINDINGS-DETAIL.md, NEEDS-VALIDATION.md and findings.json. It needs Node.js for two bundled validators, a model that supports parallel sub-agents, and an OS-enforced sandbox before it will run any of your project's code.

Good for

Maintainers who want a first-pass vulnerability review of their own repo
Security engineers who want a repeatable, evidence-first audit workflow
Developers who have a focused security question about one area of code

Install security-audit

Claude Code

Copy this prompt.

Install the security-audit skill in Claude Code. Run this command:

npx skills add https://github.com/cloudflare/security-audit-skill --skill security-audit

Then check that it loaded and tell me how to use it.

2
Paste it into Claude Code. It shows the change and asks before running it.

Requires Node.js.

From its listing on skills.sh. Source.

What's in the skill

When it loads

Security guidance and vulnerability review for codebases, APIs, services, CLI tools, libraries, and daemons. Use for security questions, focused reviews, vulnerability research, security audits, or pen tests. Run the complete workflow only for explicit codebase audit or pen-test requests, full/comprehensive/end-to-end reviews, or requested report artifacts.

Sections in SKILL.md

Operating modes
Platform terminology
Universal execution safety
Full audit setup
Write isolation
Full audit planning
Coverage and prior runs
Run profiles and scope
Core principles
Require a boundary and result
Use bounded local evidence
Respect source visibility
Separate priority from certainty
Recommend the smallest effective source fix
Full audit workflow
Anti-patterns

Key facts

Author: Cloudflare
Repository: cloudflare/security-audit-skill
Installs on skills.sh: 73,434 (2 Nov 2026)
GitHub stars (repo): 56,225 (2 Nov 2026)
License: MIT
Last commit (repo): 14 Sep 2026
What installs: 20 files: SKILL.md, 14 companion Markdown guides, report-schema.json and 4 Node .cjs validators and tests
Needs: Node.js for the validators; a model with tool use and parallel sub-agents; an OS-enforced sandbox to run target code
Output: Reports written to ~/security-audit-skill/<repo-name>/run-<N> unless you pick another folder
skills.sh audits: Gen Agent Trust Hub pass, Socket pass, Snyk warn

What it includes

Default: answers security questions without writing files
For focused reviews and methodology questions Claude uses only the relevant parts of the skill and creates no output directory.
Read · Guidance mode
Runs all six phases for an explicit audit or pen-test request
Claude resolves the target, output folder and source ref, then runs reconnaissance, hunting, validation, structured output, verification and reporting.
Try: “security audit this codebase”
Runs code · Full audit mode
Map architecture, trust boundaries and a coverage ledger
Claude writes architecture.md and coverage-ledger.json so later phases know what has been covered.
Write · 1. Reconnaissance
Isolated hunter agents work through ledger units
Each hunter gets its own scratch folder, and critic waves look for gaps.
Runs code · 2. Coverage-led hunting
A fresh verifier tries to disprove each candidate
Each unique candidate goes to a new agent that did not find it.
Runs code · 3. Candidate validation
Write confirmed, needs_validation and rejected records
Records go to findings.json and are checked against report-schema.json.
Runs code · 4. Structured output
Fresh agents verify the final source claims
Material replacements get another independent verifier.
Read · 5. Independent record verification
Derive REPORT.md, FINDINGS-DETAIL.md and NEEDS-VALIDATION.md
The report is written from the verified records and the coverage ledger.
Write · 6. Target-neutral report
quick, standard (default) or deep, plus scope and budget
quick is one hunter wave, deep splits units further. You can scope to paths or a diff and set a budget as a maximum number of agent invocations.
Read · Run profiles
Node validator for findings.json
A zero-dependency script the parent agent runs in phases 4 and 5.
Runs code · validate-findings.cjs
Node validator for coverage-ledger.json
A zero-dependency script run after the ledger is created and after each update.
Runs code · validate-coverage-ledger.cjs
14 Markdown guides Claude loads by target type
Memory safety, AI and LLM, web protocol and auth, client-side, supply chain, cloud, protocols and RPC, resource exhaustion, data isolation, desktop and mobile, plus attack classes, recon, hunting and reporting.
Read · Domain guides

From the official docs

Permissions

Medium risk

A full audit reads your whole codebase, writes report files, and may run build or test commands from the project. Use a real sandbox for any code execution and keep the report folder private.

What it can do

Read your data
Create and change things
Delete things
Read-only mode

Watch out for

Running your project's code needs a sandbox you set up
The README requires an OS-enforced sandbox with no external network, a cleaned environment and resource limits. Without it the skill keeps leads as needs_validation. Do not run untrusted code outside one. github.com
Reports list real weaknesses in your code
REPORT.md and findings.json describe vulnerabilities. Keep the output folder private and do not commit it to a public repo. github.com
It writes files, and only inside the repo if you say so
The default output folder is outside the target. The skill writes inside the repo only if you pick a folder that version control ignores. github.com

Security advisories

No published security advisories as of Nov 2, 2026.

Technical details

Turning on read-only	The skill defines no read-only switch.
Talks to	The skill says target-controlled code must run with no external network. Your AI provider receives the code and findings Claude reads
Scopes	Claude reads the whole target codebase Claude writes run files to an output folder outside the repo (default ~/security-audit-skill/<repo-name>/run-<N>) Claude runs two bundled Node validators and, inside a sandbox, build or test commands from the target

Questions

What does the Security Audit skill do?

It makes Claude audit a codebase for vulnerabilities in a fixed workflow and write a report. Findings are labelled confirmed, needs_validation or rejected, and a different agent must verify each one.

How do I run an audit?

Start Claude Code in or pointed at the codebase and ask in plain words. The README gives three examples: "security audit this codebase", "find security vulnerabilities in ./src" and "do a security review, output to ~/audits/my-project". A direct audit or pen-test request starts full audit mode. A narrower question uses guidance mode.

What does it need?

Node.js for the two validator scripts, a model that supports tool use and parallel sub-agents, and an OS-enforced sandbox for running any code from the target. The sandbox must block external networking, use a cleaned environment, enforce resource limits and allow writes only to scratch folders. Without those controls, the skill keeps the lead as needs_validation instead of running the code.

How do I install it?

Use the install box at the top of this page. The README installs it with the skills CLI and offers a global option for a user-level install.

How much will a full audit cost?

The skill does not give a price. It lets you set a budget as a maximum number of agent invocations. If the budget cannot cover the minimum, it launches no agents and asks for a larger budget, a narrower scope or a different profile.

Can I trust the results?

Treat them as leads to review. The skill separates confirmed, needs_validation and rejected records and requires independent verification. The README says: "In our test runs, a single run found roughly half of the vulnerabilities that repeated runs found in total." Runs on the same repo add to earlier ones.

Is it safe to run?

Mostly, if you give it a sandbox and read the output folder before you share it. The skill says target-controlled code may run only inside an OS-enforced sandbox with no external network, and it tells agents to use dummy data and not to probe live systems. It writes files, so check where the output goes.

Information

updated Nov 2, 2026

Repository: cloudflare/security-audit-skill
Folder: skills/security-audit

cloudflare/security-audit-skill skills.sh

Sources

Docs, issue trackers and discussions this page is built from.

Show all 2

security-audit SKILL.md · github.com
security-audit README · github.com