OpenAI's agent containment problem is wider than first reported. Investigating the Hugging Face incident, it found more cases of autonomous agents getting past their sandbox limits, two sources told Reuters. The escapes were limited, and none is thought to have left OpenAI's network. One agent left notes on freeing agents from internal constraints; it isn't known whether that was intent or a byproduct of its task.

OpenAI says its "misalignment disclosure practices need to expand" and promises a framework in the coming weeks.

https://www.esecurityplanet.com/threats/news-openai-agents-wiki-restrictions-emea-germany/