Your AI wrote the vulnerability in — 86% of the time

AI-generated code compiles, runs, and demos beautifully — and most of it cannot defend against a basic cross-site scripting attack. The model optimized for "works", not for "hostile input".

The numbers are brutal. When over 100 large language models were put through security testing, 86% produced code that failed to defend against cross-site scripting, and 88% were open to log injection. One in five referenced packages that don't exist — and attackers now register those phantom names as malware and wait for your npm install. It's called slopsquatting. Meanwhile AI-assisted commits leak secrets at roughly double the baseline rate, and a scan of 1,400 vibe-coded production apps found hundreds with exposed secrets and live PII — bank details included.

None of this shows up in your demo, because your demo only ever feeds the app friendly input. The attacker's first move is unfriendly input: a <script> tag in a name field, a newline in a log line, a package name your model hallucinated. Trust nothing that crosses your boundary.

Two checks from this doctrine

Real value, not a teaser — two of the 5 checks in full. The rest are in the one-pager.

Every input is hostile until validated, every output dangerous until escaped — the AI assumed neither.