Yes. The skill is free, with no trial, no card and no catch. You enter an email address, and the zip and the plain .md arrive as attachments. Use it on as many of your own apps as you like. It is a free one-off from authority.md, alongside our paid frameworks, Tools and Playbooks, none of which you need to buy to use it.

Will it test sites I don't own?

No. Before anything else it asks you to confirm that you own the code and the domain, or have written permission to test them. If you can't, it reviews only the code you share and makes no live requests. Where you do confirm, live checks are read-only: GET and HEAD requests to your own domain, kept to about 50, with no logins, no form submissions and no payloads.

Does it fix the problems?

It writes the fix brief, and your coding agent makes the changes. The skill does not touch your code during the audit. It ends with a paste-ready brief for Claude Code or another coding agent: a separate stage for every finding, each ending in a check and a commit you can revert on its own. When you come back after fixing, it re-checks each finding and marks it Fixed, Still open or Changed.

Is this a penetration test?

No. It is a disciplined self-audit run by your AI assistant, with read-only live checks and no exploit code. It will not call your app secure, and it states what was checked and what was not covered. Where a customer, regulator or insurer requires a penetration test, get one from a qualified professional. This skill is not a substitute for that.

Free Claude Skill

Security Audit for AI-built apps

Built your app with Claude Code, Cursor or Lovable? This skill checks the code and your live site, ranks what it finds by what it could cost you, and writes the fix brief for your coding agent.

What does it check?

Section A

Secrets

Keys printed in logs, error messages, or returned in API responses.

Section B

Things served that should not be

Paid downloads, private documents or data exports placed in a public or static folder.

Section C

Who is allowed to do what

Permission checks done only in the browser: a hidden button is not access control.

Section D

Money

Payment webhooks that do not verify the signature, or parse the body before verifying it.

Section E

Input and output

Server code that fetches a URL the user supplies, which can be pointed at internal services.

Section F

Abuse and surprise bills

Unauthenticated endpoints that call a paid AI API, turning your key into a free service for strangers.

Section G

AI features

Model output rendered as HTML or executed as code.

Section H

Headers and configuration

Debug mode, verbose stack traces, or framework version headers in production.

Section I

Dependencies and supply chain

CI workflows using unpinned third-party actions, or printing secrets to logs.

Section J

Personal data

Personal data in logs, analytics events or URLs.

Compared October 2026

How is it different from other security audit skills?

Cloudflare published an excellent open-source security audit skill in September 2026. It is built for security researchers auditing large codebases. This one is built for founders shipping apps with AI tools. Both are free.

Feature	Cloudflare security-audit-skill	authority.md Security Audit
Built for	Security research across many system types, from web apps and CI pipelines to native code and mobile	Web apps, SaaS products and stores built with AI coding tools
Checks the live site	No. Source review only, by design	Yes, with your permission: read-only checks of your own domain for exposed files, paid assets, secrets in browser code, headers and more
Runs in	Coding agents with parallel sub-agents, plus Node.js	Claude Code, Claude.ai (upload the zip), or any assistant that can read your code
Ranks findings by	Likelihood and impact, with a verdict of confirmed, needs validation or rejected	What it could cost the business: money, customer data, accounts, surprise bills
Ends with	Reports and a validated findings file	A plain-English report and a paste-ready fix brief for your coding agent, then a re-check
Catalogue	Broad attack classes	The specific mistakes AI-built apps make: public paid files, client-side prices, unsigned webhooks, open AI endpoints, prompt injection

If you're auditing a large or unusual system, use Cloudflare's. It goes deeper. The source is at github.com/cloudflare/security-audit-skill.

What does the report look like?

Situation: A founder runs a store selling downloadable digital products, built with Next.js on Vercel using an AI coding agent. Products are delivered by email after a verified Stripe webhook. The founder asks for a Standard audit before running a promotion.

With the founder's permission, one read-only request:

request

GET https://example-store.com/products/premium-pack.zip
200 OK, Content-Type: application/zip, 6,120 bytes

The paid file downloads without signing in or paying. Filenames follow the product slugs listed on the public catalogue pages, so every product URL can be guessed.

Verdict: Proven.

Severity: High. No customer data is exposed, but every paid product is free to anyone who guesses the pattern, and one shared link spreads it.

Fix: move the files out of public/ to a folder the server reads but never serves; keep delivery by email after the verified webhook; if download links are needed, issue signed links that expire.

Confirm: the same GET returns 404, and a test purchase still delivers.

What won't it do?

✦Test any system the user has not confirmed they own or are authorized to test.
✦Send anything other than read-only GET or HEAD requests, attempt logins, submit forms, or run high-volume requests against a live site.
✦Write working exploit code, or include payloads that would work against third parties. Describe the attacker's steps in plain language instead.
✦Print full secret values in the report, chat or fix brief.
✦Change the user's code itself during the audit. It writes the fix brief; changes happen in a separate, deliberate step.
✦Inflate severity, or present hardening notes as vulnerabilities.
✦Claim the app is "secure". The report states what was checked and what was found, and lists what was not covered.
✦Replace a professional penetration test where a customer, regulator or insurer requires one.

Questions

Frequently asked questions

The skill is delivered by email, as two attachments, so we need an address to send it to. That is the only reason the form asks for it. We only send you updates about new skills and tools if you tick the box, and you can leave that list at any time. If you don't tick it, you get the skill and nothing else. Details are in the privacy policy.

Another free one: Just Be British rewrites blunt AI emails into proper British ones.

Browse the frameworks See the Tools