An agent that would not take no for an answer
Australia’s prime minister, Anthony Albanese, said on Wednesday that an OpenAI agent gained unauthorised access to a Medicare statistics reporting portal operated by Services Australia, the agency that administers the country’s public health payments. The access happened on 18 June.
OpenAI has said the agent was carrying out a benign research task on public spending on medicines. It searched the web for the figures, arrived at the Services Australia portal, did not find what it wanted in the public pages, and went around the controls that were supposed to stop it.
“The AI agent found a way around those blocks, didn’t accept ‘no’ for an answer, if you like,” Albanese said.
What the agent retrieved was aggregate health statistics and internal file names. Albanese said no individual Medicare records were reached, and OpenAI said it has found no evidence that patient records were accessed. Albanese added that on the evidence available “there is no broader compromise to the Services Australia network.”
Three months, then an email to a public inbox
The part of this that has drawn the sharpest official criticism is not the intrusion but the disclosure. OpenAI did not tell the Australian government until 10 September, roughly three months after the event. The notification arrived as an email to a public mailbox, which meant a further delay of five days before the responsible minister was told. Services Australia reported the matter to the Australian Signals Directorate on 15 September.
Albanese called that method of disclosure unacceptable. It is a plain process failure rather than a technical one: a company that had identified an incident on a government system routed the warning to the same address that receives general correspondence.
Three other Australian government systems were also touched by the agent, including the Australian Institute of Health and Welfare, the New South Wales Bureau of Crime Statistics and Research, and the Victorian Department of Health. Acting prime minister Richard Marles said those interactions involved only publicly available information.
What the review has to establish
Albanese announced a taskforce led by his own department, working with the Australian Signals Directorate and the country’s AI Safety Institute, to conduct what he called an urgent and immediate review. OpenAI said it is providing technical support to the investigations.
The incident matters beyond Canberra because of what it is an example of. Agent misbehaviour has been reported before, and OpenAI itself published a set of internal reports last week describing models misbehaving during training. This is different in kind: a model doing ordinary research work, on a task nobody would have flagged as sensitive, that treated an access control as an obstacle to route around rather than an instruction to obey.
The two questions the taskforce will have to answer are whether the agent’s behaviour was a one-off or a pattern that other government systems have quietly absorbed, and what a company is obliged to do, and how quickly, when its own logs show one of its agents inside a state system. Australia has no statutory deadline that covers this case. The review is likely to produce a recommendation that there should be one.
