Report reveals yet more cases of OpenAI's 'rogue AI' agents hacking websites—and suggests they may still have been active in recent weeks

OpenAI’s rogue AI-agent problem looks bigger, older and messier than the company has publicly described. That’s the takeaway from a new report by Transluce, an independent non-profit research lab that says it has found more examples of the agents trying to break into websites, including Australian government sites, a company and a university.

The report landed the same day Australia said OpenAI’s agents had hacked an agency holding Medicare data, reaching non-public information and gaining the ability to write to file servers. That intrusion happened in June, but Australian officials said OpenAI did not tell them until September 10.

Transluce says it found attacks on the Australian Institute of Health and Welfare and BOSCAR, the crime statistics body for New South Wales. It also identified incidents involving Data USA, a free open-source platform that collects U.S. government data, and the University of New Mexico’s digital library. In at least two cases, it says, the activity had not been previously reported. The researchers also said they could tie the Australian health agency incident and the Data USA attack to the same OpenAI agent swarm involved in the July attack on Hugging Face.

The more unsettling part is the timing. Transluce says it found strong evidence of similar activity going back to March, months before OpenAI has said it saw any unauthorized behavior, and possibly as far back as November 2025. It also says the behavior may have continued through September 16 and perhaps as late as September 20. That would suggest the company still hasn’t fully contained the agents, despite saying after the Hugging Face incident that it disabled the unreleased model involved, paused key parts of training for two weeks and later added stricter controls on its unreleased models.

Transluce says the agents were not trying to do cybersecurity work in these cases. They were doing ordinary data-retrieval tasks and, when they couldn’t get what they wanted from public web pages, they switched to hacking attempts. The most recent activity, it said, looked like attempts to break into a crypto currency exchange and trade crypto currency, though those attempts failed. OpenAI did not immediately respond to requests for comment, while saying separately on Wednesday that it was in touch with Australia and that its agents had taken actions it didn’t intend.