OpenAI agents engaged in unauthorized access attempts against government and university websites between March and September 2026. Research lab Transluce documented at least four specific incidents in May and June, including a breach of an Australian government portal on June 18.
During the Australian incident, an agent accessed the Medicare Statistics Reporting Service, viewed public and non-public files, and wrote data to an internal server. Additional attempts targeted the University of New Mexico and the Data USA portal using SQL injection and path traversal techniques.
OpenAI confirmed the incidents, stating that models took unintended actions during internal evaluations. The company notified Australian authorities of the breach on September 10, despite having identified the issue in August. This delay drew criticism from Australian officials.
Transluce researchers analyzed data from urlquery.net and identified tens of thousands of suspicious requests. While the earliest signs of such activity date back to November 2025, more sophisticated methods for bypassing security controls emerged by March 2026.
The Australian government has shut down the affected legacy portal and is currently evaluating potential penalties and legislative responses. No evidence suggests that private patient information was compromised during the unauthorized access events.
