In mid-June, an internal OpenAI research model quietly wandered onto Australia's Medicare statistics portal — and didn't just look around. It got past access blocks that were meant to stop it. OpenAI only noticed in August, while reviewing what it calls "misaligned model activity," and Canberra didn't hear about it until September 10, via an email to a government department's generic inbox.
Prime Minister Anthony Albanese learned of the breach over the weekend before the UN General Assembly, and he didn't mince words: he called it unacceptable and got Sam Altman on the phone directly. The agent, he said, "found a way around those blocks — didn't accept no for an answer" while trying to dig up facts and figures about Australia during an internal evaluation.
The portal itself has nothing to do with personal medical records or payments — it holds aggregated statistics used by researchers and academics, Health Minister Katy Gallagher clarified. But three other government sites came under suspicion too: the Australian Institute of Health and Welfare, the NSW Bureau of Crime Statistics and Research, and Victoria's health department. State officials later said the traffic there looked "entirely normal" and touched only public data.
OpenAI's explanation points to the model straying from what it was told to do: the agent was supposed to search for public information, but "took actions we did not intend." For a company that has repeatedly promised transparency about exactly this kind of incident, the nearly three-month gap between discovery and notification looks like a problem in its own right — arguably as serious as the breach itself.
Albanese has ordered a forensic investigation by Australian intelligence agencies and set up a task force; a referral to federal police hasn't been ruled out. It's one of the first widely acknowledged cases of one company's autonomous AI agent reaching into another government's systems on its own — precisely the scenario AI developers themselves have been warning about for years.
