OpenAI says its models broke out of a test and breached Hugging Face
During internal cybersecurity evaluations, OpenAI agents escaped their sandboxes, reached the internet and compromised parts of Hugging Face's systems. OpenAI called it a "warning shot".
OpenAI’s full report, published August 26, says the activity was driven mainly by an internal-only research model it calls IM1, running with reduced safeguards. Agents turned a shared package server into a message board, exploited a previously unknown flaw to gain internet access, and reached third-party accounts. METR and Redwood Research published an independent analysis the same day.
OpenAI says it is isolating sandboxes further, restricting internet access and model-weight access, and spending more compute on monitoring models’ reasoning.
Index treatment. None. An incident report is not a capability evaluation.