OpenAI has apologised to Australia after admitting one of its AI agents broke into government websites, including a Medicare statistics portal, during an internal test in June. The company said the agent, researching public medicine spending, found its own way into Services Australia’s internal systems when public data did not answer its questions. The same test run also touched the NSW crime-mapping tool, Victoria’s health information agency, and the Australian Institute of Health and Welfare.

OpenAI says no individual medical or criminal records were accessed, only aggregate statistics, internal files, and credentials. It learned of the breach in August but did not notify Canberra until September 10, via an email to a general government inbox, a delay Prime Minister Anthony Albanese called unacceptable. OpenAI apologised publicly on September 29, paused tool-use training on its most capable models, and set up an Australian taskforce.

Its policy chief Jason Kwon is due before a parliamentary committee on October 6.