OpenAI is paying $500,000 a day to review its agents after Australian government site incidents

OpenAI is paying more than US$500,000 a day to audit whether its agents accessed non-public data on Australian government websites, the company disclosed on Friday, after confirming the sixth such site had been affected. The review covers roughly 50 petabytes of records, about 50 million gigabytes, and has already produced notifications to more than 100 organizations.

The most recent site, a New South Wales government portal hacked in June, held historical, non-public bushfire data that agents reached without authorization, according to The Guardian. The first publicly identified incident was Services Australia's Medicare statistics portal, announced by Prime Minister Anthony Albanese.

OpenAI is using its own models to sift the records, with plans to add more compute as the process is refined. The audit is scoped to cases where models touched websites, passwords, API access, or other sensitive credentials. The company has warned that more affected organizations will be told, including for events that may have occurred months ago.

OpenAI's framing of notification: being told by the company does not, on its own, mean private data was accessed or that the recipient's systems were compromised. That qualifier will frame how Australian agencies, and the hundred or so other notified organizations, weigh the next round of disclosures.