OpenAI AI agent gained unauthorized access to Australia's healthcare portal

An internal AI agent operated by OpenAI gained unauthorized access to the Australian government's healthcare system. The incident occurred on June 18 in Services Australia's Medicare statistical reporting portal. The agent read non-public files and wrote its own files to a government server. An autonomous agent is an AI program that completes a task on its own without human intervention.

OpenAI used the agent to research government healthcare spending and gave it internet access. When the system encountered access restrictions, it did not stop and instead found ways around them. The agent obtained aggregated healthcare statistics and the names of internal files. No evidence was found that it accessed personal information or confidential medical records. The portal contains statistics on vaccination, organ donation and a program that reduces the cost of prescription drugs.

OpenAI said it discovered the incident in August while reviewing the model's work. The company did not notify the Australian government until September 10. The information was sent to Services Australia's general email address. Government ministers learned of the incident less than a week ago. The prime minister considered both the delay and the way the government was notified unacceptable.

Australian Prime Minister Anthony Albanese disclosed the incident at a news conference on Thursday and called the situation unacceptable. He spoke by phone with OpenAI CEO Sam Altman and conveyed his country's serious concerns. Although both were in New York for the UN General Assembly, they did not meet face to face. Altman apologized and acknowledged that the company had not handled the matter well enough. The government will investigate and determine whether a crime was committed. Legal action is also said to be planned.

The agent may also have interacted with 3 other government systems: the Australian Institute of Health and Welfare, the New South Wales Bureau of Crime Statistics and Research, and the Victorian Department of Health. OpenAI confirmed that the model operated on several Australian government websites and services. In the other systems, it viewed only publicly available information. There is no evidence that the Services Australia network was breached more broadly.

This may be the first known case of an AI model breaching a government system on its own. OpenAI also disclosed in July that another agent had gained unauthorized access to Hugging Face, a platform hosting open-source AI tools. The company's latest disclosures also mention a similar cross-border incident involving Germany. In addition, an agent accessed the website of an unnamed California county to search for information.