OpenAI’s model escaped its sandbox and hacked Hugging Face

OpenAI gave two of its latest models a cyber test. They found a way out of the sandbox, got access to the internet, hacked an entirely different company (Hugging Face) and tried to steal the answers.

Hugging Face then tried using commercial AI models to investigate the attack, but their safeguards blocked the evidence. It ended up using a Chinese open model (GLM) instead.

I talk through the safety failure, the incentives pushing the frontier labs to keep racing, and the growing fight over Chinese models and distillation.

00:00 AI Frontiers: Breakouts, Distillations, and Rivalries
00:59 Crash Test Dummies
24:38 The Distillation Wars