Vibe coding · Security · Free & open source
Vibe Coded App Security Audit: Free Cloudflare Skill Explained
You shipped it in a weekend. Did anyone check the locks?
Cloudflare open-sourced a free security audit for AI coding agents. Here’s what it does, what it quietly can’t do, and why the clever part is who checks whom.
It’s 11:47 on a Sunday night. The app you built this weekend is live. You typed a few sentences, an AI wrote the rest, and somehow there’s a login page, a database, and a friend who just signed up and texted, “wait, you made this?”
You feel like a magician. Then, somewhere around your second glass of water, the other thought arrives: what exactly did I just put on the internet?
You could ask the AI that wrote it. It will answer in the same cheerful voice it used for everything else. But that’s asking an author to proofread their own essay — they read what they meant, not what’s on the page.
The short version
Cloudflare’s free, MIT-licensed security-audit skill makes your coding agent audit a project in six stages: map the app, hunt for flaws, try to disprove each one, then verify and write a report. You install it with one command and ask for a full audit. It finds candidates. It does not prove your app is safe.
What the Cloudflare security-audit skill is
Picture a building inspector who arrives with a whole crew, each person assigned one job. That’s the idea. A “skill” is a bundle of instructions you hand to a coding agent, and Cloudflare published theirs on GitHub under the MIT license. Install it, and your agent stops being the eager helper and starts playing the auditor, splitting into separate sub-agents with narrow roles.
Here’s the part worth remembering. The agent that finds a suspected flaw is not the one that confirms it. Fresh agents are brought in to try to disprove each finding, and more arrive to re-check the final claims against your source code. As we read it, that’s the whole trick: peer review for a machine, built to weed out the findings that merely sound convincing.
Why vibe coders should care about AI-generated code security
Here’s the uncomfortable bit. When you build by describing what you want, you’re shipping code you may not be able to read. That’s the whole magic of vibe coding. It’s also the whole risk.
One yardstick: in 2025, Veracode handed 80 coding tasks to more than 100 AI models and found insecure code in 45% of the outcomes. Java fared worse (over 70%), and on cross-site scripting tasks the models failed 86% of the time.
⚠️ Read that number with care
It comes from a security vendor, was scored by the vendor’s own scanner, covers 80 set tasks rather than real finished apps, and dates from July 2025. It tells you the risk is real. It doesn’t tell you how risky your app is.
How to actually run it, step by step
Fair warning before the steps: this is a menu, not a restaurant review. Everything here comes from the project’s README, and we have not run the skill ourselves yet. Treat it as a map of the instructions, not a tested walkthrough.
1. Check you have the basics
The README asks for a coding agent whose model supports tool use and parallel sub-agents, plus Node.js for the validators. It doesn’t name specific agents, so whether yours qualifies is a question for your tool’s own docs.
2. Install it
npx skills add https://github.com/cloudflare/security-audit-skill --skill security-auditAdd --global if you want it in every project instead of just one.
3. Ask for a full audit, in plain words
Wording matters. A direct request like “security audit this codebase” starts the full six-stage run. A casual “is my app secure?” is treated as a question, not an audit, and you won’t get the report files unless you ask for them.
4. Read the right file first
By default the results land in a security-audit-skill folder in your home directory, one folder per run. Open REPORT.md first (the confirmed problems), then NEEDS-VALIDATION.md (suspicions that still need a human). Fix the confirmed ones one at a time, like a to-do list you didn’t ask for.
💡 Run it more than once
The authors say a single run found roughly half of what repeated runs found in total. Later runs reuse earlier results to go after the gaps, so a second pass is the cheap way to catch more.
The honest limitations
Every good tool has a fine-print section. Here’s ours, and theirs.
⚠️ What we couldn’t confirm, and what the project itself says
We haven’t run it. Everything above comes from the README and the Veracode release. We have no results from a real app to show you.
No cost figures. The README gives no token or price information. Several agents plus repeated verification passes will almost certainly use more than a single prompt — that’s our inference, not a published number. Watch your usage on the first run.
It won’t run your code without a sandbox. Executing builds or tests needs an OS-enforced sandbox with networking off. Without one, those leads stay in “needs validation,” which means some work lands back on you.
“Not a vulnerability” is a category. The project treats weaker defense-in-depth gaps as hardening notes, not confirmed flaws. A clean REPORT.md is not a clean bill of health.
One more thing, said plainly: an audit like this reads your code. It can’t see how you set up your database, who has access to your hosting account, or what personal data you’re collecting and why. Those are easy to get wrong, and they’re yours to check by hand.
What to check by hand, either way
The skill reads code. It doesn’t walk up to your front door and rattle the handle. You can.
Secrets. No passwords or API keys inside your code or your repository. Our .env file security guide covers this.
Who can read what. Log out, open a private window, and try to reach another user’s data by changing an ID in the address.
Your database rules. Make sure “anyone can read” wasn’t left switched on from the prototype stage.
Personal data. Write down what you collect. If you can’t say why you need a field, remove it.
A plan before the build. Security is easier to ask for up front. See our vibe coding planning guide.
Who should actually care
Worth it if: you’re about to put an AI-built app in front of real users, you’re comfortable running a terminal command, and you’d rather hear about a hole from a report than from a stranger.
Maybe not if: you’re still on a throwaway prototype, or your agent doesn’t support parallel sub-agents. Start with the hand checklist above.
Skip if: you were hoping it would replace a human review of an app that handles payments or sensitive records. It won’t.
Quick answers
Is the Cloudflare security-audit skill free?
The skill itself is open source under the MIT license. What your coding agent charges you to run it isn’t stated anywhere we could find, so check your own plan’s usage.
Do I need to be a developer to use it?
You’ll need to run a terminal command and use a coding agent. Beyond that, we can’t say how beginner-friendly it feels, because we haven’t tested it. Reading the report and fixing the findings is the harder half.
Can I point it at someone else’s code?
Our advice: only audit code you own or have permission to test. Security testing someone else’s project without permission can cause real trouble, whatever tool you use.
Does a clean report mean my app is secure?
No. It means this process didn’t confirm any flaws in the code it examined. Configuration, access settings and the data you collect are outside what it checks.
The bottom line
Cloudflare’s skill is a smart, free way to get a second opinion on AI-written code, built on one good idea: whoever finds the flaw shouldn’t be the one to sign it off. It’s a solid habit before launch. It’s not a guarantee.
Your next step: pick your smallest project, install the skill, ask for a full audit, and read only REPORT.md first. Write down how long it took and what it cost you — that’s the information nobody has published yet.
