45% of AI-generated code contains a vulnerability: what the data on 380,000 apps says

🚨 They analysed 380,000 apps built with vibe coding. The results are disturbing.

The numbers: 380,000 resources scanned, over 5,000 with no protection

Researchers scanned more than 380,000 public resources created with Lovable, Replit, Base44 and Netlify.

More than 5,000 apps had no adequate protection.

Nearly 2,000 were exposing sensitive data: financial information, corporate documents, private conversations and admin access.

And it doesn’t stop there.

In a separate study, Veracode (the well-known cybersecurity company) found that 45% of AI-generated code contains at least one vulnerability traceable to the OWASP Top 10.

The scan comes from Israeli company RedAccess and was independently verified by Axios on 7 May 2026. The 45% figure comes from the 2025 GenAI Code Security Report by Veracode, which tested over 100 models across 80 coding tasks.

The four mistakes that keep repeating

The most absurd part?

These aren’t sophisticated attacks. It’s almost always the same mistakes.

Here are the main ones:

1. Secrets in the frontend

Stripe keys, API keys and credentials dropped into browser-side code.

2. No Row Level Security

Login works, but without RLS policies users can reach data they should never see.

3. Authentication mistaken for authorisation

Change the ID in the URL and you get another user’s data.

4. Debug routes left public

Endpoints like /debug, /test or /admin end up online with no protection at all (as in the Baudr case by Grenbaud a few months ago).

These are the four checks I run on every release in custom software development projects, along with authentication, permissions, and configuration.

The Moltbook case

Researchers gained full access to Moltbook’s database in a matter of minutes (remember the famous social network for AI agents?), exposing 1.5 million tokens, thousands of emails and private conversations.

All because of a misconfiguration in Supabase’s RLS policies.

The researchers were the team at Wiz, who documented the exposure on 31 January 2026: 1.5 million authentication tokens, 35,000 email addresses and private messages between agents, all reachable through a Supabase key left in the browser JavaScript.

An automated check before every push

💡 A practical tip: if you use Claude Code or Codex, ask it to create a hook that runs before every push to main, checking for exposed secrets, unprotected routes, CORS configuration, RLS policies and authorisation tests, and blocking the push when it finds a problem.

It doesn’t replace a full security review, but it keeps a lot of trivial mistakes from reaching production.