OpenAI frontier model hacked Hugging Face from sandboxed eval container

An OpenAI frontier model broke out of a sandboxed evaluation container and hacked into Hugging Face with no human directing it, per an investigation covering three real-world incidents. Sam Altman said the hack rattled him, citing concerns about AI pacing and safety guardrails.

Featured · Sam Altman