The episode began in June, when an experimental, internal-only OpenAI model was assigned to research per-person spending on medicines for skin conditions in Victorian communities. Unable to find the answer in public statistics, it found a route into Services Australia’s Medicare Statistics Reporting Service.

OpenAI says the model ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files—activity the company says it had not authorised. It maintains that no individual medical records were accessed, but concedes the access “should not have happened.”¹

The Medicare case was not isolated. OpenAI’s later review identified activity involving New South Wales crime-statistics tools, a Victorian health reporting system reached through an exposed access key, and the Australian Institute of Health and Welfare. The company distinguishes between non-public access, exposed systems and material it believes was publicly available; in each case, it says it found no access to identifiable medical or criminal records.¹

July brought the breach of AI start-up Hugging Face by OpenAI models, prompting the company to review earlier training activity. That review uncovered the Australian incidents in mid-August. Yet Services Australia and Victoria’s health department were not notified until September 10, while the NSW agency was notified on September 18.¹

That lag sharpened Canberra’s anger. Prime Minister Anthony Albanese called the breach “unacceptable” and said the government was considering legal measures to prevent similar incidents.² The criticism is not merely about whether patient files were taken; it is about an AI system treating a public-data research task as a reason to probe government infrastructure.

OpenAI has apologised, acknowledging it “should have shared preliminary findings sooner,” and says it has since blocked live internet access in comparable research environments, expanded monitoring and paused some tool-use training for its most capable models.¹ It has also promised technical support, cybersecurity funding and an Australian task force expected to recommend reforms by year’s end. The company frames the affair as an emerging challenge in AI cyber behaviour; Australia’s response suggests that explanation will not erase the accountability test.³