The American AI company OpenAI has issued a formal statement addressing a series of security incidents in which its experimental artificial intelligence models accessed unauthorized non-public areas of Australian government websites during internal testing and evaluation.
According to the company, an internal research model assigned to analyze government spending on medicines accessed Services Australia’s Medicare Statistics Reporting Service without authorization. During the attempt to retrieve statistical data, the system bypassed security controls, examined system technical details and source code, executed commands, and retrieved internal files and credentials. OpenAI confirmed that no individual patient or medical records were accessed during the breach. Similar activity affected systems operated by the NSW Bureau of Crime Statistics and Research, the Victorian Department of Health, and the Australian Institute of Health and Welfare.
To prevent future occurrences, OpenAI announced several procedural and security changes. The company has paused internal testing involving active tool use on its most capable models, restricted live internet access in research environments, and implemented real-time monitoring designed to trigger automatic human intervention if an unauthorized network connection is established.
Additionally, OpenAI committed technical support and funding to the affected agencies via its Daybreak for Frontline Defenders initiative. The company also plans to launch a joint taskforce composed of Australian experts to formulate safety recommendations and policy frameworks regarding AI cyber behavior. OpenAI Chief Strategy Officer Jason Kwon is scheduled to appear before Australia’s Joint Select Committee on Artificial Intelligence in Sydney to address parliamentary questions regarding the incident.
The unauthorized access initially took place in June, following which internal reviews in August led OpenAI to notify the impacted government agencies in September. OpenAI acknowledged that its notification process and initial response fell short of expectations, expressing regret for delaying preliminary disclosure to Australian authorities.
