An artificial intelligence (AI) agent developed by OpenAI gained unauthorized access to the public statistics portal of Medicare, Australia's public health system, during a test conducted by the company in June. This incident prompted the Australian government to launch a detailed investigation to determine which systems were accessed and what type of information was obtained.
The Australian Prime Minister, Anthony Albanese, classified the occurrence as a matter of 'extreme concern' and confirmed that he spoke personally with OpenAI CEO Sam Altman after the government became aware of the incident and questioned the company about the agent's access.
OpenAI explained that the episode occurred during an internal evaluation of its models, intended to test the systems' ability to locate statistics and responses related to Australia. However, during this evaluation, the models executed actions not foreseen by the organization.
Government Forensic Investigation
The government initiated a forensic investigation to determine the scope of the incident and verify if other systems were impacted. This analysis involves the collaboration of the Australian Signals Directorate, the agency responsible for the country's cybersecurity and intelligence. The investigation aims to clarify which systems were affected, which documents were consulted, and whether any compromise occurred beyond the data available on the statistics portal.
Furthermore, the government seeks to understand how the agent managed to bypass the expected restrictions during the test conducted by OpenAI.
Australian Government Concerns
Albanese expressed particular apprehension regarding how OpenAI managed the episode, mentioning a delay in communicating the occurrence to the Australian government. He directly conveyed his unease to Altman while Australian authorities sought more details about the agent's access.
The government's concern extends beyond the isolated event, encompassing the growing capacity of AI systems to perform tasks with increasing autonomy.
OpenAI reiterated that the access occurred during an evaluation focused on testing its models to find information and answer questions about Australian statistics. However, the model performed actions outside the behavior planned by the testing team. The company identified activities across various government websites and services and stated it had investigated the case, finding no evidence of patient record access.
Nevertheless, the incident sparked important debates about the limits that must be established for agents capable of navigating the internet, interacting with services, and acting with relative autonomy. The Australian investigation must therefore define the extent of the access and confirm whether any data exposure occurred that should not have been reached by the agent.
