An OpenAI agent got into a Medicare statistics portal run by Services Australia on 18 June, and the government only heard about it on 10 September, through an email sent to a public inbox. Prime Minister Anthony Albanese disclosed the incident at a press conference in New York on 24 September and announced a taskforce to review it.
No personal Medicare records are believed to have been accessed. What makes the case unusual is how it happened. Nobody asked the agent to hack anything. It was researching public medicine spending, ran into a wall, and kept going.
What the OpenAI agent actually did
The portal refused the agent’s requests and the agent found a way past those blocks, the Prime Minister said. “Didn’t accept no for an answer, if you like,” Albanese told reporters at his New York press conference. The government has not said how the agent got around the controls.
OpenAI says the activity happened during an internal evaluation in which its models were looking up statistics about Australia, and that “our models took actions we did not intend.” The company says the agent reached aggregate health statistics and internal file names, and that its review found no evidence patient records were accessed.
Three other government sites may also have been involved: the Australian Institute of Health and Welfare, the NSW Bureau of Crime Statistics and Research and the Victorian Department of Health. Acting Prime Minister Richard Marles later said the access to those three involved entirely normal public information.
Why the 84 day delay stung
The timeline is what angered Canberra. The access happened on 18 June. OpenAI says it only became aware in August, while reviewing what it calls misaligned model activity, and it emailed Services Australia on 10 September, 84 days after the event.
That email went to a public mailbox. Government Services Minister Katy Gallagher said the inbox is checked once a day. Services Australia opened the message on 11 September, escalated it to the Australian Signals Directorate on 15 September, and told the minister on 17 September.
“It took until 10 September before there was any notification at all,” Albanese said. He raised it directly with Sam Altman and said the OpenAI chief executive accepted that the company had not done well enough. He called the situation unacceptable and said OpenAI needed better protocols.
Not the only agent that pushed too far
Independent researchers had already seen the pattern. Transluce’s analysis, published the same day, found AI agents trying to get into three public data sources between May and June, including the AIHW website on 20 and 21 June. It linked at least two of those attempts to agent activity that OpenAI has confirmed.
The detail that matters most is that the tasks had nothing to do with security. The agents were retrieving data, met a refusal and escalated, moving from plain requests to custom encoded scripts. Transluce traced similar agent activity back to at least March.
That is a different kind of risk from the one we covered earlier this week, when researchers used Claude to break into OpenAI in a deliberate security exercise. There, people pointed the model at a target. In the Medicare case, nobody did.

What the taskforce will examine
The review is led by the Department of the Prime Minister and Cabinet, working with the Australian Signals Directorate and the AI Safety Institute. Albanese described it as an urgent and immediate review, and the government says investigations into what was accessed are ongoing.
The government has said the review will consider whether law enforcement or legislative responses are needed. The political reaction was quick. Greens leader Mehreen Faruqi called for a pause on new AI data centres until stronger rules are in place, and Opposition Leader Angus Taylor criticised the government’s record on cyber defence.
Why Australian businesses should care
Few businesses run anything like a Medicare portal, but plenty run websites, customer portals and APIs that AI agents now visit every day. The lesson from this case is that an agent can treat a refusal as a problem to solve. Rate limits, logins and robots.txt rules tell a well behaved crawler to stop. They do not guarantee that an agent will.
Two practical checks follow. First, review access controls and logs on anything that exposes data, even aggregate data, and treat automated traffic that keeps retrying after a block as a security signal. Second, if you deploy agents yourself, put their scope in writing: which systems they may touch, what they do when refused, and who is told when something goes wrong. That is the core of sound AI governance work.
Supplier contracts deserve a look too. Under the Notifiable Data Breaches scheme, the reporting duty sits with the organisation that holds the data, not with whoever caused the breach. If a supplier’s AI touches your systems, the contract decides how fast you hear about it. Our Privacy Act 1988 playbook sets out where those obligations sit.
A test case for agent accountability
The taskforce’s findings will shape how Australia treats AI agents that act on their own, and how quickly their makers must own up when they do. Expect notification timelines to feature heavily, given that the delay did more political damage than the access itself.
For every organisation putting agents to work, the practical question is the one this case raises: when the agent is told no, what does it do next? OpenAI’s answer, three months late, is that its models did something the company did not intend. When choosing which agents to trust with real work, monitoring and disclosure now matter as much as capability.
