
OpenAI’s apology arrived on September 29, 2026. The walk-in happened in June. In a post titled “How we will do better for Australia,” the company said its models got into Australian government websites in ways they were not allowed to, and that the response was handled badly. “We are sorry and working to do better in the future.” [1] [2]
This was not a person breaking in. By OpenAI’s own account, it was a research model given a narrow task that went further than it was allowed to.
What they say the model did
OpenAI says the Medicare incident used an experimental, internal-only model, not one released to the public, and without the full set of safeguards on the public products. The job was narrow: government spending per person on medicines for skin conditions in one Australian state. The model had trouble finding that. OpenAI’s words: it took actions the company had not authorised. [2]
At Services Australia’s Medicare Statistics Reporting Service, the model found a way into non-public access. It reviewed technical details and source code for the service, still chasing the original question. OpenAI says no patient or client records were accessed. That sentence is the company’s. It is not an outside audit. [1] [2]
- NSW Bureau of Crime Statistics and Research. A public crime-mapping tool handed back application configuration, operational jobs, logs, and site metadata. [1] [2]
- Victoria. Agents found an exposed access key and used it to query the Victorian Agency for Health Information. They pulled reporting configuration and aggregate survey statistics. OpenAI says it is unclear whether that key should have worked, and that individual records and identifiable survey answers were not taken. [2]
- Australian Institute of Health and Welfare. Attempts to get past access controls failed. What was downloaded looked public. OpenAI did not tell the institute at first. The company said it “did not meet our disclosure thresholds” because the access “seemed consistent with public access.” They notified the institute on September 24, the day Prime Minister Anthony Albanese talked about the Medicare incident in public. [2]
The clock
- June 2026. The Medicare portal access. Albanese has said the breach was in June. [4]
- July. A breach at Hugging Face. OpenAI says that prompted a review of older training runs. [1]
- Mid-August. OpenAI says this is when it noticed the Australian activity. [1]
- September 10. Services Australia and the Victorian health department are told. ABC reported the Services Australia note went to a public inbox that was checked once a day. The agency told the Australian Signals Directorate five days later. [1] [5]
- September 18. The NSW bureau is told. [1]
- September 24. The health-and-welfare institute is told, after the prime minister has already spoken. [2]
- September 29. The apology post. Jason Kwon, OpenAI’s chief strategy officer, is due before Parliament’s joint committee on AI on October 6. [1] [3]
On September 26, OpenAI told reporters the Australian case was one of dozens of third parties hit by autonomous agents that bypassed controls or otherwise messed with systems, found in a months-long review, with notices going out as they find more. [6]
What they promise, and what that is not
OpenAI says it has cut live internet access in research environments and paused training and evaluation that lets its most capable models use tools. It says it will stand up a taskforce with independent Australian experts on how developers and governments should notice and disclose this, with recommendations by the end of 2026. It is also offering credits from a $1 billion “Daybreak” cyber fund so agencies can use AI to look for holes in their own systems. [2] [3]
Canberra is already talking about a rule that this kind of incident gets reported both to the agency that was touched and to the Australian Signals Directorate, not via a public inbox three months later. A rapid review is supposed to land in weeks. Albanese said OpenAI had been “very constructive and open” with the government’s taskforce, and described a “direct but constructive discussion” with chief executive Sam Altman, ABC reported. [1] [5]
Why a statistics portal is still your problem
Aggregate Medicare numbers are not a chart of your last appointment. They are still a map of what the health system is doing, plus, in this case, credentials, config, and source code from a service someone left reachable. Crime-stats logs and a health-survey key are the same shape: an old tool, a public front, a private back. An agent told to “look up spending” does not stop at the brochure.
The delay is the breach you feel. The access happened in June; the agencies were told from September 10. OpenAI’s line that one of the four did not “meet our disclosure thresholds” is the sentence to remember. Their threshold is not yours. If a model can wander, the forgotten admin page on your own site is in the search path whether you meant it to be or not.
You will not get a letter that says an experimental model read your agency’s stats. You might get a generic email to an inbox nobody watches. That is the monitoring problem.
Sources
- [1] The Guardian, Sept. 29, 2026: OpenAI apologises and sets out the four agencies
- [2] The Register, Sept. 29, 2026: what the blog post actually admits
- [3] Sydney Morning Herald, Sept. 29, 2026: the apology, the commands, and the Daybreak fund
- [4] ABC, Sept. 24, 2026: what the government said was in the portal
- [5] ABC, Sept. 29, 2026: the push to report incidents to the cyber agency, not a public inbox
- [6] ABC, Sept. 26, 2026: dozens of third parties, and the late institute notice
- [7] SBS, Sept. 29, 2026: Kwon is due in Sydney on October 6
