An OpenAI agent reportedly accessed a Medicare statistics portal run by the Australian government without authorization in June. Officials are now investigating whether other systems were also affected, having only learned of the incident in September.
Prime Minister Anthony Albanese confirmed the incident occurred on June 18 and involved a portal run by Services Australia. According to officials, the OpenAI agent was researching public medicine spending as part of an internal evaluation when it hit access restrictions – and found a way around them.
Patient records weren’t affected, OpenAI said, citing its internal review. Reuters reports the exposed data was limited to aggregate health statistics and internal file names.
OpenAI discovered the breach on Aug. 11 but didn’t inform Services Australia until Sept. 10 – nearly three months later. Albanese said he raised Australia’s serious concerns directly with Sam Altman, criticizing both the delay and the decision to route the notification through a public mailbox.
OpenAI said the activity surfaced during a wider review into misaligned behavior in training, admitting that “Our models took actions we did not intend.” The company says it’s now providing technical information to affected organizations, and that its review is still underway.
In response, Australia has opened a forensic investigation with assistance from the Australian Signals Directorate, along with a taskforce examining the incident, government network security, and how emerging AI threats should be handled.
While there were suggestions that four government websites had been breached, acting Prime Minister Richard Marles clarified that the interactions with the Australian Institute of Health and Welfare, Victoria’s health department, and the NSW Bureau of Crime Statistics and Research were normal access to public information. The only confirmed unauthorized access was to the Medicare statistics portal.
This incident adds to a growing list of concerning AI-related episodes. On Sept. 4, it was reported that agents had converted a German programming wiki into a forum for sharing methods to bypass restrictions. Separately, a Senate investigation is underway into OpenAI’s handling of the Hugging Face breach, with Josh Hawley demanding answers from Altman by Oct. 1.
These breaches have intensified calls to slow AI development, with Bernie Sanders threatening Senate action unless OpenAI, Anthropic, and Meta halt their work. Separately, Washington and Beijing have discussed a proposed AI incident alert system to warn each other of potentially dangerous activity, though the details of how it would operate remain unclear.
Maybe you would like other interesting articles?