Is your vibe-coded app secure?

This is for you if you built your own app

You had an idea, and instead of waiting months for a developer, you built it yourself with Lovable, Bolt, Replit, Cursor, Claude Code or Codex. That takes ingenuity and nerve. Most owners never try.

Now that app holds things worth protecting: the time and money you put into it, the income it brings in, and the trust of every customer who typed their details into it.

The people looking for holes just got better tools

In April 2026, Anthropic, the company behind Claude, announced an AI model called Claude Mythos. It found thousands of security holes nobody knew about, in every major operating system and web browser. Anthropic chose not to release it to the public because of what it could do in the wrong hands.

Not every AI company holds back. In February 2026, Anthropic reported that three AI labs, DeepSeek among them, had copied Claude’s abilities at industrial scale, and warned that copied models “lack necessary safeguards.”

The skill of finding holes in software is spreading into tools with nothing to stop someone pointing them at your app. Someone who wants your customers’ data no longer needs to be an expert. They need an app nobody has checked.

It has already happened

• Moltbook, early 2026. Its founder said publicly that he didn’t write one line of code for it. Security researchers at Wiz got into its database in under three minutes and found 1.5 million access keys and thousands of private messages open to anyone who looked.
• Tea, July 2025. A dating safety app left a storage area open. About 72,000 images leaked, including roughly 13,000 selfies and photo IDs that users had sent to prove who they were. A class action lawsuit was filed within days.

Neither owner set out to put their customers at risk. Nobody had read the code with an attacker’s eyes before the attackers did.

Who is reading your code

I’m Atiba de Souza. I was a software architect before I was a marketer, and I’ve been running AI inside my own companies since about 2014.

In January 2024 I stood on Ryan Deiss’s stage and told the room that every business would soon have AI agents doing real work inside it. That is where we are now. Ryan has since said he learned more about AI from me than from anyone else.

So if you went and learned to build with AI, good on you. You did something most owners never will.

What you learned isn’t wrong. It’s incomplete.

The tools take shortcuts. In the middle of a fix, Claude, Codex or whatever you use makes a quick choice that works today and isn’t safe for the long run. It never tells you it made that choice, so you never know.

It happens to us too. A client found a bug in one of our apps and we fixed it fast. In that fix, Claude had the browser talk straight to the database. Because we build with checks in place, we caught it within hours and closed it. Most owners have no such checks, so nobody would ever notice.

Dr. Lisa Faast and Jason Spaven built their app on The Architect, the build structure we teach, which gives an app a sound start. Our audit still found drift. I can only imagine how much drift there is in an app that started without that structure.

That is why we built this audit. You keep building. We read everything your AI wrote and show you what it did that you never saw.

Atiba de Souza
Impossible Means Go

Who has used it

Dr. Lisa Faast and Jason Spaven

Dr. Lisa Faast, CEO of DiversifyRx, and her build partner Jason Spaven had us audit the app they built, at our Las Vegas event in October 2026. They had started from a sound structure, and the audit still found problems. They used the report to fix their app.

What the review built into Claude Code and Codex actually checks

You may already run the review your AI tool comes with. Here is what each one is built to look at, in its maker’s own documentation.

• Claude Code’s /code-review checks the changes on your branch and looks for bugs. It runs inside the session where you are working, following the same project instructions that session follows. You can point it at a folder, but only if you know which folder to ask about.
• Claude Code’s /security-review is told to look only at security problems your new changes add, not the ones already in the app.
• Codex’s /review checks your uncommitted changes, one commit, or a branch against the one it came from. On GitHub, it reviews the changes in a pull request and flags only the two highest priority levels.
• OpenAI’s Codex Security is a separate tool that scans connected GitHub repositories for weak spots. Its cloud version is in research preview and needs a Codex cloud environment set up first.

A review of what just changed never looks again at the code from your first month of building, which is where the oldest shortcuts sit.

What our audit reads

• All of the code, including what was written months ago.
• The project’s full history. A password or key you deleted from a file is still stored in the history, and still works until you change it.
• Who can reach whose data: whether one customer can see another customer’s records.
• Nineteen areas in all, from sign-in and payments-adjacent data to what happens when your AI features are fed hostile text. Every area gets a line in your report: what we found, “checked, clean”, or “not checked” and why.
• Every serious finding names the file and the line, and that line is checked again to make sure it really exists before the report goes out.
• Two AI readers, working separately. Each reads the whole app without seeing the other’s notes, and the report marks every problem both of them found.
• A person reviews the report before it is sent. Nobody at our company opens your code.

We use Anthropic’s AI, the company behind Claude, to do the reading. What changes is the job: a separate read of your whole app, against a written list of areas, with people reviewing the report.

What happens to your code

• We only read your code. We never change it, never run it, and never sell or publish it.
• To read it, your code is sent to Anthropic’s AI through their business service. Anthropic’s commercial terms say: “Anthropic may not train models on Customer Content from Services.”
• Under its standard terms, Anthropic deletes what we send within 30 days. It keeps something longer only if it is flagged as breaking its usage policy, or if the law requires it.
• Our own copy of your code is deleted when your report is done. Your report shows the time it was deleted.
• Our GitHub app can only read, only the repo you choose. You can remove it any time in GitHub, under Settings, Applications, Installed GitHub Apps.
• Nobody at our company opens your code. Only the AI reads it. People review the report about it before it is sent.
• Your report quotes the lines it talks about, so it holds short pieces of your code. It never repeats a password, a key or a customer’s personal details: it names the file and tells you what to change. The report link stops working after 30 days, or sooner if you ask.