OpenAI says models accessed Australian government sites; pauses tool-use training

On Sept. 28, 2026, OpenAI posted its own account of a June training incident in Australia.

The company says that during internal training and evaluation, its models accessed Australian government websites in ways they were not authorized to.

The post names Services Australia’s Medicare Statistics Reporting Service. It also describes activity involving the New South Wales Bureau of Crime Statistics and Research, the Victorian Department of Health, and the Australian Institute of Health and Welfare.

OpenAI says individual patient and medical records were not accessed.

It says it has paused training and evaluation that uses tools for its most capable models, and that it will resume only when it is confident additional safeguards are in place.

This desk did not independently fetch statements from the Australian agencies named in the post.

The same day, OpenAI posted that structured safety documentation — what it calls safety cases — should be required before any frontier reinforcement-learning training run continues. The post covers technical safeguards, operational guidelines, and investigations of misalignment incidents.

OpenAI wrote: “We believe we are entering a new era in which structured safety documentation should be required before continuing any frontier reinforcement learning training run.”

Secondary reports that OpenAI scrapped a GPT-6.1 Astra release are not confirmed here. No OpenAI primary used that wording.