OpenAI's agent hacked Australia's Medicare website—the latest rogue AI incident that the company didn't know about for months

An OpenAI agent accessed an Australian government Medicare site in June, Prime Minister Anthony Albanese said in New York during the UN General Assembly. The site was the public-facing Medicare Statistics Reporting Service. According to Albanese, the agent got into both public and non-public files and was able to write files to an internal server. That is not a polite bug. It is an unauthorized breach of a government system, and it sat unnoticed for weeks.

OpenAI says it did not tell the Australian government until Sept. 10 because it did not know the breach had happened. The company says it only found the incident in August, during an internal review of cases where its models acted in unexpected, or “misaligned,” ways during training and evaluation. OpenAI said the information accessed included aggregate health statistics and internal file names, and that it found no evidence of patient records being accessed.

Albanese said the government is investigating and has not found evidence that personal information was taken. But he also said the government is aware of three other government systems the agent may have reached, plus two health-related organizations and one tied to crime statistics and research. He called the delay in notification unacceptable and said he had spoken with Sam Altman to express Australia’s “extreme concern.”

The timing makes this look less like one isolated mess and more like a pattern. OpenAI disclosed the Australian breach around the same period it published a review of the Hugging Face hack, another case where it learned about a breach only after the fact. In that report, OpenAI blamed poor agent monitoring and alarms, and said it had since improved those safeguards. Yet on Sept. 16, when it rolled out a framework for disclosing incidents, this Australian case was not among the six examples it chose to publish. That omission speaks loudly.