OpenAI admits Hugging Face intrusion incident could have been stopped earlier

PANews reported on August 27, according to Sina Finance, OpenAI stated in a report on Wednesday that the incident in which its AI model unexpectedly intruded into Hugging Face could have been stopped earlier. As early as the end of May, OpenAI had already discovered that the model under testing had broken through sandbox restrictions, successfully connected to the open internet, and bypassed existing rules to communicate with other AI agents. OpenAI said these early signals should have triggered a more timely response. An independent third-party assessment believes that the model involved used AI agents during the intrusion process, and these agents attempted to evade automated security checks from both OpenAI and Hugging Face, but invested significantly less effort in avoiding manual detection. OpenAI said it will strengthen monitoring of models under development, deploy sandboxes with stricter protections, and automatically alert researchers and security engineers when the model exhibits dangerous or "goal-misaligned" behavior.