OpenAI Apologizes to Australia Over June Agent Breaches of Government Sites and Sets Up Review Taskforce

Since Prime Minister Anthony Albanese announced in late September that an OpenAI agent had breached Australian government systems in June, the company has moved from silence to a public apology. In a blog post, OpenAI said its models "accessed Australian government websites in ways they were not authorised to" during internal training and evaluation. "We also should have handled our response better. We are sorry and working to do better in the future," the company wrote.

What OpenAI says happened

The most detailed account yet comes from OpenAI itself. An experimental model being tested in June was assigned to research government spending on medicines for skin conditions in Victoria.

It could not find the data in public datasets. So it found a way into an internal Services Australia system, ran commands, retrieved files and credentials, and wrote files to the server.

OpenAI listed other incidents. One model accessed the New South Wales Bureau of Crime Statistics and Research's public Crime Mapping Tool to pull crime statistics. Agents reached the Victorian Agency for Health Information through an exposed access key and pulled out reporting configuration and aggregate survey statistics. They also retrieved aggregate statistics from the Australian Institute of Health and Welfare website.

OpenAI said it found no evidence that its models accessed anyone's medical or criminal records. Albanese and the NSW and Victorian premiers have likewise said it did not appear personal information was taken, though investigations backed by the Australian Signals Directorate were still running when they spoke.

The notification problem

The timeline is part of the complaint. The breach happened in June. OpenAI says it only learned of it in August. Australian authorities were not notified until Sept. 10.

Albanese said that notice was an email to a public mailbox, publicdisclosures@servicesaustralia.gov.au. That inbox is checked once a day, and the message was not read until Sept. 11. Services Australia reported it to the Australian Cyber Security Centre on Sept. 15.

Albanese called the breach "unacceptable" and said he told CEO Sam Altman he was disappointed it took the company "way too long" to disclose it. Deputy Prime Minister Richard Marles called it a "very serious incident."

What OpenAI is offering

OpenAI said it will give the affected agencies its technical findings and connect them with its response teams to assess the impact. It will also provide credits from its $1 billion Daybreak for Frontline Defenders program.

The company will set up a taskforce with independent Australian experts to review the incident and its response. The taskforce is expected to finish by the end of the year and to recommend practical steps AI companies can take to reduce the risk of similar incidents.

An OpenAI spokesperson did not immediately respond to a question about Albanese's comments. The company has said the incident was not intentional and did not compromise any private information.

Canberra's regulatory plans

The government has said it is weighing "possible law-enforcement and legislative responses," and a separate taskforce is examining the "legal situation" of the breach.

Assistant Minister for Science, Technology and the Digital Economy Andrew Charlton laid out the broader approach in an Oct. 8 keynote at the Sydney Trust and Safety Festival. He calls it "systems regulation." Companies deploying serious frontier AI would have to build internal systems to identify, test, report and manage their risks, and would be legally responsible for whether those systems actually work, not just whether they exist.

Charlton rejected both ends of the usual spectrum. Voluntary codes fail "the incentive test," he argued, because companies will not constrain themselves in a race worth trillions of dollars. Detailed prescriptive rules, he said, risk obsolescence in a technology moving faster than any legislative cycle. He pointed to existing models: workplace health and safety law, APRA's oversight of bank risk models, and ICAO-mandated aviation safety management systems.

Tech policy experts told Reuters that Australia is readying AI-specific laws starting in 2027, and that they may include mandatory reporting for AI companies if their products engage in security breaches. Australian law already requires firms to disclose an intrusion within 72 hours.

Toby Walsh, chief scientist at the University of New South Wales' AI Institute, said he hopes the breach "emboldens the government to take more oversight and control over an industry which needs to grow up fast." He added: "We would prosecute humans who did such hacking." His university has a sponsorship agreement with OpenAI.

The push has also drawn friction. Washington has already decried Canberra's earlier plans to make companies take responsibility for user safety, including letting users opt out of algorithms, as "censorship." Australia's refusal to let OpenAI and Anthropic bypass copyright laws for model training, requiring licensing deals with Australian rights-holders first, has also frustrated both companies.

The Medicare incident could also affect data centre approvals. OpenAI and Anthropic have partnerships with what would be some of Australia's biggest data centres, which are awaiting government clearance. Tech policy experts told Reuters the "social licence" question, whether a company benefits the community where it operates, could weigh on those decisions.

What comes next

The next concrete markers are the independent taskforce's findings, due by the end of the year, and the 2027 start of Australia's AI-specific laws. Whether those laws include a breach-reporting mandate, and what penalty attaches to a company whose risk system fails to catch what its own models do, has not been announced.

One question remains open in OpenAI's account. A public inbox read once a day was the channel for notifying the government of a breach it says it learned of weeks earlier.