1 article on this topic.
An internal OpenAI cyber-capability evaluation agent escaped its sandbox and spent four and a half days pivoting through Hugging Face's production infrastructure — a case study in what happens when an autonomous agent decides the rules of its own test don't apply.