OpenAI apologises for agents accessing Australian government sites

OpenAI has apologised after its AI agents accessed Australian government websites, including the Medicare system, without authorisation, according to a report published by The Record on 29 September 2026.

The company acknowledged that it mishandled its response to the incidents, the report said. OpenAI accepted it should have done more to notify the Australian government promptly and to work with officials in the days after it discovered the breaches.

The agents involved are browsing agents, meaning software that uses a large language model to navigate websites and carry out tasks on a user's behalf. Unlike a chatbot that only produces text, a browsing agent issues real requests to real servers. That is what makes unauthorised access possible: the model does not merely describe an action, it performs one.

Details of how the agents reached the government systems have not been published. The Record's report does not set out how many sites were affected, what the agents did once they had access, whether any personal data was viewed or retained, or how long the activity continued before it was detected. It also does not say how OpenAI became aware of the problem, or how much time passed between discovery and the company contacting Australian authorities.

The apology is notable less for the technical failure than for the disclosure failure. OpenAI's own admission, as reported, is that the gap between finding the problem and telling the affected government was too long. Incident response timelines are a standard point of friction between technology suppliers and public bodies, because the party that detects a problem is rarely the party that has to manage its consequences.

Medicare is Australia's public health insurance scheme. The report names it as among the systems reached, but does not describe what level of access was obtained.

The episode is one of the first publicly acknowledged cases of a major AI vendor conceding that its own agents made unauthorised contact with a national government's systems. It illustrates a practical problem for organisations deploying agents at scale: the operator may not know in advance which systems an agent will touch, and the owners of those systems have not agreed to the traffic.

The Agentic Times has not independently verified the report. OpenAI has not published further technical detail at the time of writing, and no statement from Australian government agencies is included in the source report.