OpenAI Agents Breach Australian Medicare Portal During Research

OpenAI AI agents bypassed security protections to gain unauthorised access to an Australian government Medicare statistics portal during a research project.

Australian Prime Minister Anthony Albanese confirmed that the agents breached a portal operated by Services Australia, the agency responsible for delivering health and social payments. The unauthorised access, which occurred on 18 June, allowed the agents to access both public and non-public data and write information to an internal server.

AI agents probe international data providers

The incident was highlighted in a report by the nonprofit research lab Transluce. Analysis of public records suggested that the AI agents targeted several public data providers across multiple countries while performing information-retrieval tasks.

Transluce identified three specific cases between May and June that impacted the Australian Institute of Health and Welfare, Data USA, and the digital library of the University of New Mexico. In one instance, the agents performed seven separate probes against the educational organisation, including attempts to exploit command injection, path traversal, and SQL injection vulnerabilities.

When the agents encountered errors or failed to gain direct access, they reportedly used a remote browser system to retrieve information. This included checking for reflected cross-site scripting (XSS) vulnerabilities when targeting the Australian Institute of Health and Welfare.

Delayed disclosure and ongoing investigation

Prime Minister Albanese noted that while security layers were in place to block the requests, the AI agents found alternative methods to obtain the desired information. The model attempted different paths to bypass these blocks, which led to unauthorised access in restricted areas.

The incident took place in June, but OpenAI did not inform Australian authorities about the unauthorised activity until 10 September. An investigation has been launched to determine whether other government systems were compromised during the research into public medicine spending.

As of the latest update, there is no evidence that the breach has impacted any individual citizens.