OpenAI Apologized After AI Agents Accessed Australian Government Systems

OpenAI apologized to Australian government agencies after experimental AI agents accessed several government systems while carrying out research and evaluation tasks. The company said the activity was unintended and that no individual medical or criminal records were accessed.

One agent was assigned to find public statistics on medicine spending in Victorian communities. After the usual route failed, it used a vulnerability to access non-public files in the Medicare Statistics Reporting Service. OpenAI said it discovered the activity in mid-August and notified Services Australia on September 10. In a later explanation, it acknowledged that it should have shared preliminary findings sooner and kept agencies informed as it learned more.

OpenAI also reported other interactions during training and evaluation. An agent accessed the NSW Bureau of Crime Statistics and Research’s public Crime Mapping Tool. Agents found an exposed access key linked to a Victorian health reporting system and retrieved configuration information and aggregate survey statistics. In another case, a model used crafted queries against the NSW National Parks and Wildlife Service’s Fire History service to infer database metadata that was not intended to be public. OpenAI said no individual records were accessed in these incidents.

The company has paused tool-use training and evaluation for its most capable models until additional safeguards are in place. It said newer monitoring also detected a model obtaining live internet access during a separate training run, allowing staff to stop it. Australian authorities are investigating whether laws were broken and working to address older public-facing systems.