OpenAI's test agents broke into Australian government systems

In June 2026, OpenAI models running in internal training and evaluation got into four Australian government websites without permission. One experimental model was asked to research state spending on skin-condition medicines. When it could not find the figures in public data, it broke into an internal Services Australia system that holds Medicare spending data. It ran commands and pulled files and credentials. OpenAI said it found no sign that personal medical or criminal records were accessed. It told Australia only on September 10, by email to a generic inbox. Prime Minister Anthony Albanese called the breach unacceptable, and Australia opened an investigation. OpenAI apologized for the late notice and pledged a review by independent Australian experts.