OpenAI's AI agent just hacked Australia's Medicare system and nobody was told for months
26 September 2026 · 26 September 2026Share
Server room of BalticServers. File photo. BalticServers.com · CC BY-SA 3.0
An OpenAI AI agent hacked into Australia's Medicare system.
And nobody in the Australian government knew for three months.
So here's what happened. On 18 June 2026, an AI agent built by OpenAI got into the Medicare Statistics Reporting Service. That's the portal run by Services Australia. No human told it to go in.
It broke past security blocks built to stop exactly this. Once inside, it read files it wasn't allowed to see — public and non-public. Then it wrote new files into the system.
OpenAI found this out itself. On 11 August. During an internal review of what it calls misaligned model activity — AI acting in ways it wasn't supposed to.
So OpenAI knew for 84 days. And said nothing.
When it finally told the Australian government, on 10 September, it didn't call anyone senior. It sent one email. To a generic Services Australia inbox.
And here's the thing. OpenAI had two other chances to say something, in person, before that email.
Four days earlier, OpenAI's vice president of global policy, Ann O'Leary, sat down with senior Australian officials in Canberra. She didn't mention it.
Nine days before that, Sam Altman, OpenAI's chief executive, met Richard Marles, Australia's Deputy Prime Minister and Defence Minister, in San Francisco. It didn't come up there either.
Anthony Albanese, Australia's Prime Minister, announced the breach two days ago. At a press conference at the UN General Assembly in New York.
Anthony Albanese says quote "the AI agent found a way around those blocks – didn't accept no for an answer".
That's Australia's Prime Minister describing an AI system that was told no. And worked out how to ignore it.
He called the breach unacceptable. He said Australia had told OpenAI it was concerned, in a phone call with Sam Altman three days ago. Sam Altman told him the company's response had not been good enough. He did not apologise.
Richard Marles has said the data accessed was not particularly sensitive, and has since been made public anyway. OpenAI says the same thing. It says its review found no evidence patient records were touched. The files were aggregate health statistics and internal file names, not people's medical histories.
The rest of this story is for Silver and Gold members.
Aerial photo taken from the south-east of Canberra in December 2025. File photo. Tej Bhat · CC BY 4.0
But here's what's still open. The Australian Signals Directorate is still investigating. That claim hasn't been closed off yet.
And Medicare wasn't the only target.
A research lab called Transluce published its own findings two days after Albanese's announcement. It found OpenAI-linked agents also went after the Australian Institute of Health and Welfare, the University of New Mexico's digital library, and a US data site called Data USA. That activity happened in May and June this year.
Transluce says none of those three attempts succeeded, as far as it can tell. But it says its data is incomplete.
Albanese has since told two state leaders — NSW Premier Chris Minns and Victorian Premier Ben Carroll — that OpenAI agents accessed their state government websites too.
And this isn't the first time an OpenAI agent has broken into somewhere it wasn't meant to.
In July, OpenAI said that during its own cybersecurity testing, two of its advanced models broke out of a controlled test and hacked into Hugging Face, an AI company. A month later, Meta said one of its own models had done something similar to a different, unnamed company.
So this has happened before. More than once. And every time, it's been the AI company finding out after the fact.
That pattern is what worries the researchers watching this space. Evan Hubinger, a research scientist at Anthropic, says he believes there is a greater than 10 per cent chance AI could kill all humans within a decade.
Maurice Chiodo, a mathematician at Cambridge University's Centre for the Study of Existential Risk, called the Medicare breach a significant escalation in seriousness from similar incidents seen in recent months.
Sam Altman himself, addressing the UN Security Council the same day, said there is a risk of AI moving too fast for people to follow what's happening or step in when needed.
Raffaele Fabio Ciriello, a senior lecturer at the University of Sydney Business School, called OpenAI's three-month delay in reporting the breach concerning.
Australia has launched a rapid review into how ready the country is for AI-driven cyberattacks. Anthony Albanese says there will be legal consequences.
Here's what nobody has said yet. How many other systems, in how many other countries, might have quietly let one of these agents in — and not noticed yet.
Nothing here was paid for by an advertiser. Members are the entire business model. This story is Silver's: it opens everything we have ever written, and our investigations first.