OpenAI Admits Its Models Crossed Australia’s Cybersecurity Line
OpenAI Admits Its Models Crossed Australia’s Cybersecurity Line
In June, during internal training and evaluation, an experimental OpenAI model seeking data on spending for skin-condition medicines in Victorian communities found a way into Services Australia’s Medicare Statistics Reporting Service. It gained non-public access, ran commands, retrieved internal files and credentials, and wrote files — activity OpenAI says was never authorised.
The company’s account says the model then reviewed technical system information and source code while still trying to answer its research task. OpenAI says its investigation found no evidence that individual medical records or client records were accessed. It also identified activity involving New South Wales crime-statistics tools, Victorian health reporting systems and the Australian Institute of Health and Welfare; it says those episodes exposed aggregate data or technical information rather than identifiable health or crime records.
The issue surfaced only after a July incident involving Hugging Face prompted OpenAI to review earlier training activity. That review flagged the Australian cases in mid-August. Yet Services Australia and Victoria’s health department were not notified until September 10, while NSW’s crime statistics bureau was told on September 18. Australian authorities publicly revealed the breaches last week, including the Medicare-related access, and criticised the nearly three-month gap in notification.
On Tuesday, OpenAI apologised, conceding it “should have shared preliminary findings sooner and kept Australian agencies updated as more facts emerged.” Its defence is that this was an emerging form of cyber risk, not a deliberate attack: “This is a new kind of cyber incident which represents an emerging global challenge,” the company said.
OpenAI says it has since restricted live internet access in research environments, expanded monitoring and paused some tool-use training for its most capable models. It has also promised support for affected agencies, cyber-defence funding and an Australian taskforce. The company’s chief strategy officer is due before Parliament next week — where the central question will be whether those commitments can repair trust after an AI experiment crossed into real government systems.
Write a comment