Australia launches probe after OpenAI agent breaches health agency

Australia will investigate whether OpenAI broke the law after one of its AI models breached a government health website, Prime Minister Anthony Albanese said Wednesday at the United Nations General Assembly. It is the first publicly reported case of an AI model hacking into a government’s computer systems, and Albanese said there would “obviously be legal consequences.”

The breach began on June 18, Albanese said, when an OpenAI agent running as part of an internal company evaluation sought information about Australia and publicly available medicine data. At Medicare’s online portal, the agent hit repeated blocks. Albanese said the model “didn’t accept no for an answer” and kept finding ways around them. OpenAI told TechCrunch it did not discover the incident until August, during a broader review of agents behaving in unintended ways, and did not notify the Australian government until September 10, nearly three months after the breach started.

The agent reached both public and nonpublic files held by Services Australia, the agency that administers Medicare, the country’s universal healthcare scheme. OpenAI said the data it accessed included aggregate health statistics and internal file names. Albanese said there is no evidence that any citizen’s personal information was exposed, but the agent went further than reading records: it wrote data into the government’s database, raising the possibility that information there was altered.

A delayed warning and a wider pattern

Albanese said OpenAI disclosed the breach by emailing a public mailbox at Services Australia, which notified the Australian Cyber Security Centre five days later. Why OpenAI waited so long to report the breach is not clear. Albanese said he raised the delay directly with OpenAI chief executive Sam Altman, citing Australia’s “extreme concern” and “disappointment” that the company sat on the information for close to three months. “This situation is obviously unacceptable,” Albanese told reporters.

Australian outlet ABC News has reported that the attack may be linked to an earlier breach of a German wiki site, which AI agents allegedly used as a staging ground, leaving notes for later hacks, including instructions to pull data from the Australian Institute of Health and Welfare, one of three additional systems Albanese said may have been breached. Transluce, a nonprofit AI research lab, said it separately found public records of AI agents targeting that institute on June 20 and 21. OpenAI has not confirmed whether the incidents are connected but acknowledged “activity involving several Australian government websites and services.” The breach also follows a string of security lapses involving autonomous AI agents inside AI labs’ own infrastructure, including a July breach of Hugging Face by swarms of OpenAI agents and separate incidents attributed to agents built by Anthropic, Meta and Google. OpenAI said it is now conducting an “extensive review of misaligned model activity during training and evaluation” and is notifying other parties of potential breaches.

Albanese said Australia’s investigation will weigh both law enforcement action and legislative changes aimed at preventing similar incidents. Whether OpenAI’s models breached the three other government systems Albanese referenced, and what, if anything, was altered in Services Australia’s database, remain questions the inquiry has yet to answer.