Australia has ordered a rapid review of whether its laws and government processes are ready for cyber incidents involving AI, after an AI agent from OpenAI got into restricted parts of a government website. Prime Minister Anthony Albanese announced the review on Sept. 24, Australian time, while in New York for the United Nations General Assembly, and said his government knew of no earlier case like it. The review's recommendations, including what AI companies must report to the government, will feed into the AI safety standards Australia plans to set in law.
On June 18, OpenAI researchers gave an internal AI model an internet research task on public spending on medicines. The model was working as an AI agent, AI that can carry out tasks on its own, such as using websites, instead of only answering questions. It ran into repeated blocks on a statistics portal for Medicare, Australia's health care system, run by the government agency Services Australia, and found a way around them. Albanese said it reached both public and nonpublic files and, according to Services Australia, wrote files to an internal server. He called it a research project that "got into areas that it shouldn't have." OpenAI says the model ran commands and retrieved internal files, credentials and aggregate statistics, but no individual patient or client records.
Government Services Minister Katy Gallagher said the portal is a standalone public website, used mostly by researchers for aggregate data, with no link to Medicare claims, payments or individual information. It holds nonsensitive figures such as Medicare spending, and officials believe no personal information was accessed. Acting Prime Minister Richard Marles called the impact relatively minor but the incident very serious. Albanese said systems at three other bodies may also have been affected: the Australian Institute of Health and Welfare, the New South Wales Bureau of Crime Statistics and Research and the Victorian Department of Health. OpenAI said on Sept. 28 that its models had retrieved technical details or aggregate statistics from all three, and that no individual records were accessed.
Australian leaders have criticized how slowly OpenAI told the government. OpenAI found the activity on Aug. 11, during a review of what its models had done. It did not tell Services Australia until Sept. 10, in an email to an inbox that researchers use to report security weaknesses. On Sept. 15, Services Australia alerted the Australian Signals Directorate, the agency that investigates cyber breaches, which is aiding a forensic investigation. Albanese said he told OpenAI CEO Sam Altman that both the delay and the way the company gave notice were unacceptable. He said Altman "clearly accepted that the company had not done good enough." Marles said OpenAI had been "working with us very cooperatively."
The review is led by the Department of the Prime Minister and Cabinet, working with the National Cyber Security Coordinator, the Australian Signals Directorate, the Australian AI Safety Institute and Services Australia. The government has asked the review for recommendations on reporting rules for AI-driven cyber incidents, notice and cooperation duties for AI companies, stronger defenses for government systems, and whether current offenses and penalties are enough. Albanese said the government would seek urgent advice on whether any offense occurred and whether to refer the case to the Australian Federal Police. He said the incident would also go to Parliament's Joint Select Committee on Artificial Intelligence.
Andrew Charlton, the assistant minister for science, technology and the digital economy, said the government wants to introduce laws setting standards for AI safety and data center construction by the end of 2026. He said it hopes to pass them in early 2027. Incident reports must be timely, full and sent to the right place, he said, and OpenAI's "fell short." Charlton said the government is also reviewing whether the law needs to change to recognize an incident carried out by an AI agent rather than directly by a person or a company. Independent Sen. David Pocock said it was "a bit rich" for the government to seek global cooperation after it had "shelved plans for an AI safety act" at home.
OpenAI apologized on Sept. 28, saying it should have shared its early findings sooner. It says it has since blocked live internet access in the research environments involved and added monitoring that would flag such activity for urgent human review. OpenAI said its chief strategy officer, Jason Kwon, would answer questions from the Joint Select Committee in Sydney on Oct. 6.
On Oct. 4, OpenAI said its review had found another case from June: a model researching wildfire statistics sent specially written requests to the New South Wales National Parks and Wildlife Service's fire-history map and worked out database details that were not meant to be public. OpenAI said it found no sign the model retrieved personal information. It is not yet known what else OpenAI's models reached, whether any Australian law was broken or what the new laws will ask of AI companies.