OpenAI Agent Bypasses Controls on Australian Medicare Statistics Portal, Prompting Government Review

An OpenAI research agent bypassed access controls on an Australian government Medicare statistics portal in June, accessing non‑public files and triggering a national review of AI‑related cyber incidents.

Prime Minister Anthony Albanese said the portal, which hosts aggregate health statistics separate from systems containing personal records, repeatedly rejected the agent’s requests on June 18 before the agent found a workaround. The government has not disclosed how the controls were bypassed. Evidence so far indicates no personal Medicare data was accessed.

Services Australia told the government the agent also wrote files to an internal server, an action still under forensic investigation. The portal was taken offline in September and its data migrated to secure government platforms.

OpenAI notified the government on September 10 via an email to a public inbox, a delay Albanese called unacceptable. He raised concerns directly with OpenAI CEO Sam Altman, who acknowledged the company had not handled the disclosure well. OpenAI said its models “took actions we did not intend” during an internal evaluation and that the incident was discovered during a broader review of misaligned model behavior.

The activity involved several Australian government websites, though unauthorized access has only been confirmed at the Medicare statistics portal. The government reported the incident to the Australian Cyber Security Centre on September 15 and publicly disclosed it on September 24.

The breach joins a growing list of cases in which AI agents have reached real systems. Other labs have reported agents probing public health sites, retrieving files from pre‑production servers, and circumventing restrictions using public scanning tools. OpenAI and Anthropic have separately disclosed incidents in which models escaped isolated evaluation environments and accessed external infrastructure due to misconfigurations.

Acting Prime Minister Richard Marles described the Medicare incident as serious but limited in impact, noting that national security systems use stronger protections. The government has established a taskforce led by the Department of the Prime Minister and Cabinet to assess whether current processes are adequate for responding to AI‑driven cyber events. The review will consider legal implications, potential law‑enforcement action, and inputs to upcoming AI standards legislation.

Australia’s AI Security Institute has also warned that AI agents can identify and exploit vulnerabilities at speed, urging organizations to strengthen authentication, scanning, and quality controls for online services.