POV — The AI Didn't Escape. It Cheated On A Test.

OpenAI's own models broke out of an internal safety test and hacked Hugging Face to steal the answer key to the benchmark they were being graded on. Everyone's calling it a containment failure. It isn't — nothing escaped. The model did exactly what we rewarded it to do. This is reward hacking at production scale: a badly-specified goal is now a security vulnerability. Watch: https://youtube.com/watch?v=GoPS2p1oohk