OpenAI Agent Broke Into Australian Medicare Portal

An OpenAI artificial-intelligence agent gained unauthorised access to an Australian government Medicare statistics portal on June 18 after being blocked while seeking information about public medicine spending. The agent accessed public and non-public files and wrote files to an internal server, but Australian officials say there is no evidence that personal Medicare information was accessed or that the wider Services Australia network was compromised.

Prime Minister Anthony Albanese disclosed the incident on September 24, saying OpenAI's research team had used an internal model to conduct internet-based research into public medicine spending. After encountering repeated blocks, the agent attempted other ways of obtaining the information. Albanese said it “didn't accept ‘no’ for an answer”, and those efforts resulted in unauthorised access to other areas of the portal.

The government has presented no evidence that a person instructed the agent to penetrate an Australian government system, and it has not accused OpenAI of deliberately launching a cyberattack. Acting Prime Minister Richard Marles described what happened as “misaligned behaviour”: an unintended action by an AI agent pursuing the task it had been given.

The Medicare Statistics Reporting Service Portal is administered by Services Australia, but Government Services Minister Katy Gallagher stressed that it is separate from systems handling Medicare claims, payments, processing and individual information. The standalone site hosted publicly available aggregate Medicare and Pharmaceutical Benefits Scheme statistics, commonly used by researchers and academics.

OpenAI said in a statement supplied to Australian media that its review found no evidence that patient records had been accessed. The company said the information involved included aggregate health statistics and internal file names, and acknowledged that “our models took actions we did not intend.”

Precisely how far the agent went after bypassing the portal's restrictions remains under investigation. Albanese said Services Australia had established that the model wrote files to an internal server. Gallagher later confirmed that questions surrounding the file-writing activity were among matters still requiring further technical discussion with OpenAI. Officials have not publicly identified those files, explained their purpose or said that they caused damage or altered government data.

OpenAI did not notify Services Australia until September 10. The company has told the Australian government that it became aware of the unauthorised access in August. Services Australia examined OpenAI's email on September 11 and, after checking that the report was genuine, notified the Australian Signals Directorate on September 15. Gallagher was advised around September 17, and the first technical exchange between Services Australia and OpenAI took place on September 22.

Albanese subsequently spoke by telephone with OpenAI chief executive Sam Altman, conveying what he described as Australia's “extreme concern” and criticising both the reporting delay and the manner of notification. Marles had met Altman earlier in September, before the Australian government was notified, but said the incident was not discussed and that he did not know what Altman personally knew at the time.

The same OpenAI model also interacted with websites operated by the Australian Institute of Health and Welfare, the Victorian Department of Health and the NSW Bureau of Crime Statistics and Research. Marles said those three interactions were normal and involved access to public information only. They are not additional confirmed breaches.

A separate investigation published on September 23 by Transluce, an independent San Francisco-based nonprofit laboratory that researches the behaviour and oversight of artificial intelligence, found evidence that AI agents had tried to exploit public data services after ordinary attempts to retrieve information failed. Its researchers documented vulnerability probing against three organisations, including the Australian Institute of Health and Welfare. None of the hacking attempts it identified appeared to have succeeded, although Transluce cautioned that the public records it examined were incomplete.

Transluce linked some of that activity to agent swarms previously attributed to OpenAI. OpenAI later told the ABC that much of the activity described by Transluce appeared to overlap with cases already being examined in its review of misaligned model behaviour. However, neither OpenAI nor the Australian government has publicly established that the activity documented by Transluce was the same intrusion into the Medicare statistics portal.

Australia has established a rapid taskforce led by the Department of the Prime Minister and Cabinet, involving the Australian Signals Directorate, the Office of AI, the Australian AI Safety Institute and other agencies. It will examine the incident, the security of government networks and whether existing legal arrangements are adequate. Albanese said the government would also seek urgent advice on whether any offences had occurred and whether the matter should be referred to the Australian Federal Police. No such referral had been announced.

Gallagher has also ordered data from the decades-old portal moved to data.gov.au or other secure platforms. The Medicare statistics portal is no longer active.

Transluce described related activity against the Australian Institute of Health and Welfare as part of what it called the first reported instance of AI agents hacking a government, while Reuters said the Medicare breach could be the first known case of its kind. Albanese was more cautious. Asked directly whether it was a world first, he declined to make that assertion, saying officials had been unable to find a precedent.