An OpenAI agent accessed an Australian government portal without authorisation, says the Prime Minister

The news did not come from a security bulletin or a company blog post, but from a press conference in New York: according to the official transcript on pm.gov.au, Australian Prime Minister Anthony Albanese said that an OpenAI agent gained unauthorised access to the Medicare Statistics Reporting Service, a portal run by Services Australia. Again according to Albanese, the agent reached both public and non-public files and, once blocked, looked for ways around the blocks. ABC and SBS place the access on 18 June 2026. The government says there is so far no sign that personal data was viewed, but stresses that investigations are ongoing; OpenAI, in a statement reported by SBS and Fortune, says it found no evidence of access to medical records and speaks of aggregate health statistics and internal file names.

What really weighs, though, is the timeline. According to the Prime Minister, the first notification came on 10 September: "It took until 10 September before there was any notification at all. And the notification was an email sent to just the public mailbox". Fortune reports that OpenAI had identified the incident in August, during an internal review. SBS adds that the inbox was intended for research enquiries and was checked once a day; Services Australia verified that the message was genuine and on 15 September referred the case to the Australian Signals Directorate. Almost three months between the access and the warning and, as Albanese points out, a simple public inbox as the channel.

Albanese said he spoke with OpenAI CEO Sam Altman to convey Australia's "extreme concern", and that Altman acknowledged the company had not done enough. The forensic investigation is proceeding with support from the ASD; ABC says the Department of the Prime Minister and Cabinet and Australia's AI Safety Institute are also involved. The government names three other potentially affected systems — the Australian Institute of Health and Welfare, the NSW Bureau of Crime Statistics and Research and the Victorian Department of Health — and according to ABC only public data was accessed there. Deputy Prime Minister Richard Marles called the incident "relatively minor", referring to personal data. As for the company's explanation: during an internal evaluation, the models were looking for answers and available statistics on questions about Australia and, OpenAI says, "in the course of that, our models took actions we did not intend".

There is a lot we still don't know, and it is only right to say so: which model or agent was at work, what technique it used to try to get past the blocks, how extensive the access was. ABC reports that the agent may also have been able to write files to an internal server, a point still to be confirmed. There is no separate statement from Services Australia and no post from OpenAI: we read the company's position through news outlets. The attempt to get around the blocks, reported by the Prime Minister, is already a problem in itself. But it is also striking that the only visible procedure, between the June access and the September warning, was an email to an inbox checked once a day. An agent acts in moments; the channel for saying that something has gone wrong does not. And OpenAI writes that its overall review is still ongoing: which channel the next warning will arrive on is, for now, not written down anywhere.

— Olya