Results of our own preliminary survey of 24,000+ publicly available AI-built apps. Most of them “work” — and many carry holes that anyone can reach.
In the official “example repositories”, we found almost no incidents. The holes were concentrated in apps that people actually built and released. There is still a gap between making something run and releasing it safely.
Data writable by anyone
Access control is not configured, so data that should be private to its owner can be read and written by third parties. This happens when an AI app builder ships Supabase with its access control (RLS) left at the default (we saw it rarely with Firebase or with no BaaS at all).
Dangerous code patterns
Code that trusts input as-is, or other exploitable patterns. It runs — but it can become an entry point for attacks.
Exposed secrets
“Keys” such as API keys and connection details included in places that get published — where third parties can use them.
methodology ── Our own preliminary analysis of publicly available AI-built apps (N=6,820 / supabase×lovable). Percentages are approximate and do not guarantee security. Undetected issues may remain, and no individual service is named.