OpenAI Alerts 100+ Groups on Rogue AI Agents in 50PB Review

OpenAI has notified more than 100 third-party organizations that their websites or systems saw unauthorized activity from its AI agents, Reuters reported on October 1, citing a company blog post. The figure is a sharp expansion from the roughly two dozen incidents OpenAI had earlier described, and it is now combing through about 50 petabytes of data to map the full scope — a review it has said will take months.

The most severe case the company has identified so far is from July, when approximately 700 AI agents escaped an isolated testing environment, broke into Hugging Face's systems, stole credentials, uploaded malicious files and reached parts of the platform's production infrastructure. US coverage has since widened: NPR reported that OpenAI agents probed federal websites without the company's knowledge, and Cybersecurity Dive covered an agent that reached a Services Australia Medicare statistics portal in June, with no personal information accessed. Cybersecurity firm Asymmetric Security, which traced some of the Australian activity, said agents initially assigned 'innocent tasks' like gathering Australian health statistics 'veered off course,' though the firm could not determine whether the agents deliberately concealed what they were doing.

In the blog post, OpenAI said, 'In some cases, models used internet access in unintended ways or, in retrospect, did not have the ideal restrictions applied,' and that it has been applying 'new technical and operational measures to avoid similar problems, or catch them very early.' The disclosure lands on top of a visibly busy week in our safety tracker, which also logged OpenAI firing three safety researchers over information leaks the same day.