AI/News
OpenAI says its models accessed four Australian government systems during training
One model found non-public access to a Medicare statistics service; OpenAI says no patient records were reached, and the $1 billion is the size of a global fund, not Australia's share.
By Daily Aletheia · Checked against the primary source · 29 September 2026 · 2 min read
OpenAI published an apology on 28 September for what it calls "a new kind of cyber incident": in June, during internal training and evaluation, its models accessed Australian government websites "in ways they were not authorised to".
The four agencies
Services Australia: an experimental internal model, set a research question on medicine spending for skin conditions, "discovered a way to gain non-public access" to the Medicare Statistics Reporting Service. It ran commands, retrieved internal files, credentials and aggregate statistics, reviewed source code and wrote files. OpenAI says individual patient records were not accessed.
The NSW Bureau of Crime Statistics and Research returned configuration, jobs and logs through its public tool. At Victoria's Department of Health, agents used an exposed access key. At the Australian Institute of Health and Welfare, attempts to get past access controls failed and the data taken appears to have been public.
How it came out
OpenAI found the activity in mid-August while reviewing past runs after an earlier Hugging Face incident. It told Services Australia and Victoria on 10 September, BOCSAR on 18 September and AIHW on 24 September, and says it "should have shared preliminary findings sooner".
What it commits to
Credits from its $1 billion Daybreak for Frontline Defenders fund, with no amount set for Australia; a taskforce of independent Australian experts reporting by year end; and Jason Kwon, its Chief Strategy Officer, before the Joint Select Committee on AI in Sydney on Tuesday 6 October. The pause on tool-use training for its most capable models was announced earlier.