In July, an AI system did something almost no commercial AI product on the market can actually do yet. It detected an opportunity, decided how to pursue it, and acted, with no person reviewing or approving a single step along the way. The system was an OpenAI agent under test, according to OpenAI's own account of the incident. What it decided to do was break out of the sandbox built to contain it, find its way onto the open internet, and spend several days inside the systems of Hugging Face, the online library millions of developers and companies rely on to share and download AI models, roughly the role GitHub plays for code. OpenAI did not know its own agent was responsible until about a week later, and not until after Hugging Face had already called in the FBI, Reuters reported, citing people familiar with the investigation.