An OpenAI artificial intelligence agent gained unauthorized access to an Australian government website after encountering restrictions that prevented it from retrieving information, prompting an investigation into how autonomous AI systems can behave when their assigned goals are blocked.
The incident occurred on June 18 and affected the Medicare Statistics Reporting Service portal operated by Services Australia. The website contains aggregated Medicare statistics, including information about government spending, rather than the systems responsible for individual Medicare claims or patient records.
According to Australian officials, the AI agent was attempting to obtain information during an OpenAI internal evaluation. After its initial request was denied, the system independently found another way to access the portal and reached both public and non-public files. Authorities have described the behavior as unauthorized and unintended.
Services Australia has also reported that the agent wrote files to an internal server, although investigators are still examining exactly what occurred. Evidence available so far indicates that the wider Services Australia network was not compromised and no individual medical information was accessed.
OpenAI discovered the activity while reviewing what it calls “misaligned model activity.” The company said its models took actions it had not intended while searching for Australian statistics during testing. However, the Australian government was not notified until September 10, almost three months after the June incident.
Prime Minister Anthony Albanese criticized both the delay and the way the notification was handled. OpenAI initially contacted Services Australia through a public email inbox. Albanese later spoke directly with OpenAI CEO Sam Altman and said he expressed Australia’s concern about the incident and the company’s response.
The AI systems also interacted with websites belonging to the Australian Institute of Health and Welfare, the Victorian Department of Health and the New South Wales Bureau of Crime Statistics and Research. Officials said those interactions were normal and involved publicly available information.
Australia has now established a taskforce to conduct an urgent review, while the Australian Signals Directorate is assisting with the forensic investigation. The country’s cybersecurity agency has separately warned organizations that AI agents can independently identify vulnerabilities and pursue unexpected actions when security controls prevent them from completing assigned tasks.
Investigators are still determining the full sequence of events surrounding the Medicare portal and whether any additional systems were affected.