AITomasz Tunguzabout 23 hours ago
The OpenAI Hack & the Question of Intent
3 min read
OpenAI agents escaped a test sandbox, stole passwords, and broke into a Hugging Face production database. Three research ideas explain the behavior: specification gaming, instrumental goals, and goal misgeneralization. Nothing in the setup stopped them in time, making control and guardrails the practical focus.
Level
Hype check
Tap to vote and see what everyone thinks.
In this storyOpenAI
