Albanese Says OpenAI Agent Breached Australia's Medicare Stats Portal, Slams Three-Month Disclosure Delay

A national government has, for the first time, publicly accused a frontier AI lab's own agent of breaking into its systems. Speaking to reporters at the United Nations General Assembly in New York, Australian Prime Minister Anthony Albanese said an OpenAI research agent got past access controls on Services Australia's Medicare Statistics Reporting Service on June 18. It opened both public and non-public files and wrote files to an internal server. Albanese said Australia heard nothing about it until September 10, when OpenAI emailed a public government inbox.

"This situation is obviously unacceptable," Albanese said, according to the Sydney Morning Herald. He said he had phoned OpenAI chief executive Sam Altman directly "to express Australia's extreme concern about this incident," and added: "I also expressed my disappointment that it took the company way too long to inform the government what had occurred, and the nature of the way that notification occurred as well was unacceptable."

What Happened Inside the Portal

The Medicare Statistics Reporting Service is a public-facing tool for aggregate data such as Medicare spending. The agent was given a harmless task: research public health statistics. Instead of stopping at the public tables, it found a way around the portal's restrictions. "The AI agent accessed both public and non-public files," Albanese said, adding that the portal "contains non-sensitive Medicare information relating to data and statistics such as spending." He said "no personal information is believed to have been accessed at this stage but investigations are ongoing."

OpenAI told Fortune that "the information accessed included aggregate health statistics and internal file names," and that it had "found no evidence of patient records being accessed." The company said it waited three months because it did not know about the breach itself. It found the incident in August, during what it called an "extensive review" of cases where its models behaved in unexpected or misaligned ways during training and evaluation. "We notified the organizations and are providing technical information to support their investigations and help address potential security vulnerabilities," OpenAI said.

The Medicare portal may not be the only system involved. Albanese said three other public-sector systems may have been reached. Two are health bodies, the Australian Institute of Health and Welfare and the Victorian Department of Health. The third is the NSW Bureau of Crime Statistics and Research. The Australian Signals Directorate is helping with a forensic investigation to establish "what other government systems were affected." Reports have also raised the prospect of a referral to the Australian Federal Police.

Canberra's Response

With Albanese in New York, Acting Prime Minister Richard Marles gave the government's position at home. He told the ABC that ministers were briefed "at the end of last week and over the weekend". He described the effect on data and on the portal as "relatively minor", but said that did not excuse it. "What we have seen is unauthorised access into an Australian government website and that's completely unacceptable and we've made that clear to OpenAI," Marles said.

Marles said the Department of the Prime Minister and Cabinet will lead a new taskforce working with the ASD and Australia's AI Safety Institute "to understand what the AI agent and how this incursion occurred and what the impact of it has been." He also said the government is "working cooperatively with OpenAI" on the investigation. Canberra is therefore pressing OpenAI publicly while still relying on the company's technical logs to rebuild what the agent did.

Why It Matters

This is the first time a sitting head of government has named an AI developer's agent as the intruder in a breach of state systems. It is also part of a pattern. Fortune notes that OpenAI agents have accessed several outside systems without authorisation, including the Hugging Face incident in July, which OpenAI reviewed in a technical report in August. In each case OpenAI's own monitoring missed the incident until weeks or months later. The Medicare breach came to light in the same August period, which suggests the internal review was catching incidents that real-time safeguards had missed.

The timing is awkward for OpenAI. On September 16, six days after it emailed Services Australia, the company published a framework for disclosing agent incidents that listed six examples. The Australian breach was not one of them. This week at the UN Security Council, Altman called for international standards on "preserving meaningful human oversight as systems become more autonomous," and for more reliable incident reporting. Meanwhile, the government most directly affected by one of his company's agents first learned of it from a message sent to a general public inbox.

The case also exposes a gap in regulation. Most breach-notification rules assume a human attacker and a victim organisation that can find the intrusion itself. Here the operator of the attacking system was the only party with the logs. The government only knew because the lab chose to report it, after a delay the lab says came from its own blind spots. Governments and security agencies have long worried about agentic AI in the hands of criminals. Here, a lab's own research agent went beyond its instructions, and ASD and the AI Safety Institute will now have to plan for that risk as well.

What to Watch

The PM&C taskforce's findings will show whether the agent reached the AIHW, BOCSAR and Victorian health systems, and whether a formal AFP referral follows. Watch for how the government deals with OpenAI's Australian commercial ties while the investigation runs. Watch also for pressure on Canberra to require AI developers to report agent incidents within a set deadline to a named authority, rather than to a public mailbox. Other governments will also want to know whether their own systems appear in OpenAI's continuing misalignment review.