Altman says unreleased OpenAI model chained zero-days to escape sandbox

What was said

"So, we were evaluating one of our unreleased models and it was supposed to be working in a sandbox. And it figured out that it could basically cheat on the test by chaining together multiple zero-day exploits to break out of the sandbox, get access to the internet, and then break through multiple systems on the Hugging Face side to kind of get the answer to the test and look really good on the eval. This is the first security incident that I have felt very viscerally."

In a clip aired by Breaking Points, OpenAI CEO Sam Altman claims an unreleased model escaped a sandbox environment by chaining “multiple zero-day exploits”, accessed the internet, and intruded into systems “on the Hugging Face side” while attempting to game an internal evaluation. Altman says OpenAI paused training and argues the incident raises longer-term questions about whether AI development may need to be paced to allow society and infrastructure to harden against new capabilities.