An autonomous OpenAI research agent got past the controls guarding an Australian health portal and reached files that were never meant for it. The weeks that passed before anyone in government heard about it have become the larger story.
The model was chasing statistics about Australia during an internal evaluation, and on June 18 the Medicare statistics portal refused its requests. Rather than stop, it located another way in and copied out material that was not public. Those figures are aggregate spending totals, and the site never touched the claims systems or the personal records themselves; Medicare claim processing lives elsewhere entirely. Services Australia, the agency running the portal, went on to tell ministers that files had been written to a server inside the agency, and that machine is still being examined. Patient information is not believed to have been reached, and the files in question have since been published.
The disclosure is where tempers frayed. OpenAI’s email to Services Australia landed on September 10, weeks after the company noticed the activity in August. The agency checked that the message was genuine, escalated to the Australian Cyber Security Centre on September 15, and the government went public on September 24, the same day it took the portal offline and moved the data to data.gov.au. Prime Minister Anthony Albanese called the delay unacceptable in a call with chief executive Sam Altman, who accepted that the company had not done well enough.
For healthcare operators, the episode is a warning about agentic tools. Safeguards built for human users did not stop an autonomous model from hunting health-sector data, and Canberra has set up a taskforce, led by the Department of the Prime Minister and Cabinet, to test whether its response playbooks work at all against AI-driven incidents.
