Original research · July 2026

The state of vibe-coded security

We ran the Merge Risk scanner over 150 public apps built with AI coding tools. Most were fine. But a stubborn share ship the one mistake that ends a weekend project: a working key, sitting in the code, readable by anyone.

Most pass. The tail is the problem.

83% earned an A. That's the honest headline, and it's good news — AI tools are not leaking secrets in the average project. But security isn't an average. One committed admin key is a full breach, and 12 of 150 repos (8%) earned a straight F.

What the AI leaves behind

The pattern is consistent. The code works, the app runs, the demo is clean — and a secret rides along, because nothing in the build process stops it. The usual suspects:

• A live key in a committed file. Stripe, OpenAI, a database URL with credentials — pasted in to make it work, then never moved to an environment variable.
• The admin key in the wrong place. A Supabase service_role key that ignores every access rule, wired into client code where the browser can read it.
• The half-fix. A secret deleted from a file in a later commit — but still sitting in the git history, still valid, one clone away.

One number we won't give you

You'll see security tools claim huge percentages of “wide-open databases.” We could have. We didn't, because our live database probe only fires when a repo commits a usable connection URL and public key together — and most don't, so we simply couldn't test most of them. A low number there means rarely testable, not proven safe. We'd rather report 14% we can stand behind than a scary one we can't.