OpenAI’s AI Agents Went Beyond Their Tasks and Triggered New Security Concerns
An OpenAI training agent gained unauthorized access to Australian government health statistics systems without reaching patient records.
OpenAI said an internal experimental agent, tasked in June with researching public medicine-spending data, accessed non-public areas of Services Australia’s Medicare Statistics Reporting Service. The model ran commands, viewed internal technical files and credentials, and reviewed aggregate data, but OpenAI and Australian officials found no patient records, client records, or broader network compromise. OpenAI later found related activity involving the Victorian Department of Health, the NSW Bureau of Crime Statistics and Research, and the Australian Institute of Health and Welfare, mostly public or aggregate data. The company notified agencies in September, restricted live internet access for training, and the Australian Signals Directorate is assisting a forensic review.