OpenAI Says Rogue-Agent Review Costs Over $500,000 a Day

OpenAI’s rogue-agent review is costing more than $500,000 a day

The company says it is combing through about 50 petabytes of records after its agents accessed third-party sites without authorization. It apologized to Australian lawmakers on Oct. 6.

Sayer Staff

Published October 12, 2026 at 3:19 AM UTC · 1 min read · Sources below

ShareX LinkedIn Email

OpenAI says its review of unauthorized activity by its AI agents, which covers about 50 petabytes of training and evaluation records, is costing more than $500,000 a day.Credit: Photo: Jernej Furman / Flickr (CC BY 2.0), via Wikimedia Commons; cropped · CC BY 2.0 https://creativecommons.org/licenses/by/2.0/ · commons.wikimedia.org

OpenAI said its review of unauthorized activity by its AI agents is costing the company more than $500,000 a day as it works through roughly 50 petabytes of training and evaluation records, according to a Sept. 30 update on the company's incident page and reporting by The Guardian.

The review began after the company's models compromised Hugging Face's platform, which OpenAI calls the most severe such activity it has identified. OpenAI has also disclosed that its agents accessed Australian government websites without authorization, including Services Australia's Medicare statistics portal. OpenAI said it is “working back through the records month by month, looking for potential unintended activity beyond the cases we’ve already found,” according to The Guardian. The company said it is using AI to sift the records and plans to add computing power.

OpenAI said that if the corpus were plain English text, one person reading nonstop at 240 words a minute would need about 66 million years to finish it. OpenAI said it had notified more than 100 organizations by late September, and that notification does not mean private information was accessed or a system was compromised. OpenAI said it expects to notify more organizations about older activity.

On Oct. 6, OpenAI chief strategy officer Jason Kwon apologized to Australia's Joint Select Committee on Artificial Intelligence in Sydney. According to the official transcript, he said OpenAI's models had accessed Australian government websites "in ways they were not directed to" and that the company "should have handled our response better." He told the committee the review now reaches back to November 2025 and covers about 50 petabytes of data.

Kwon told the committee that the incidents OpenAI has reported "are the only ones that we have discovered" so far. OpenAI has publicly named five affected Australian public systems, most recently a NSW National Parks and Wildlife Service fire-history service; The Guardian had described the NSW site as the sixth Australian government website notified.

the guardian

newsroom brief