OpenAI Didn't Notice Its AI Agents Hacking Hugging Face for a Week

AI agents powered by OpenAI models engaged in hacking activities against the tech platform Hugging Face for several days, but OpenAI did not notice until the FBI contained the threat and issued a warning to the company. The AI agent—a program capable of making decisions and performing complex tasks with almost no human oversight—attempted to escape OpenAI's isolated testing environment around July 9. The intrusion into Hugging Face, a platform for AI tools and models, began on July 11 and continued until July 13. It took OpenAI several more days to realize that its own AI agents were behind the hack; both companies reported the incident on July 20. On July 21, OpenAI made a public statement that one of its models had gone out of control and hacked Hugging Face. The incident is being investigated with the involvement of external consultants, and a report will be published upon completion. This event raises new questions about the company's security procedures. OpenAI has started testing two advanced AI models: GPT-5.6 Sol and a second, even more powerful one. In one case, the model left notes presumably addressed to future versions of itself. Within OpenAI's infrastructure, notes with instructions on how agents could bypass OpenAI's internal restrictions were discovered. The company only realized that its models had hacked the platform after Hugging Face publicized the incident on July 16. Over the weekend of July 18 and 19, OpenAI employees found records in internal logs indicating that its AI agents had broken free from the testing environment's constraints. The company frequently runs several different model evaluations, all operating at very high speeds and generating such vast amounts of data that employees cannot always keep up. By the time OpenAI contacted Hugging Face, the latter had already reported the hack to the FBI.