Australian PM Anthony Albanese said on Wednesday that an OpenAI agent gained unauthorised access to the Medicare Statistics Reporting Service portal in June. According to OpenAI, the agent did so while researching Australia’s public medicine spending to answer internal testing questions.
Authorities and OpenAI have both stated there is currently no evidence that sensitive or personal medical information was accessed.
The incident marks another significant real-world example of an autonomous AI escaping its testing boundaries to interact with external systems. As more developers report these unintentional hacks, the debate over AI regulations and disclosure requirements grows.
The development follows Capacity’s July coverage of an OpenAI model that independently hacked AI development platform Hugging Face during internal testing. Following the news, researchers highlighted concerns around increasingly capable AI agents pursuing objectives through unauthorised actions – or through methods their developers haven’t anticipated.
According to the Australian government, the OpenAI agent accessed both public and non-public files hosted on the Medicare statistics portal on 18 June. Albanese said the AI system found ways around barriers that initially prevented access, describing the situation as “obviously unacceptable”.
OpenAI said the breach wasn’t spotted until 11 August, when a review was made on their models’ activity. The company acknowledged that its models “took actions we did not intend,” but nevertheless collected non-public aggregate information from an Australian website containing non-sensitive Medicare data.
The timeline of OpenAI’s disclosure has been a large sticking point for the UN’s Australian delegation.
Australian officials have criticised OpenAI after learning the company identified the activity in August but did not notify Services Australia until 10 September. The disclosure was reportedly sent to an inbox monitored by government researchers and academics, rather than directly to senior officials or cybersecurity teams.
Albanese said he expressed “extreme concern” and disappointment directly to OpenAI CEO Sam Altman over both the delay and the method of notification.
The Australian government has since established a taskforce to investigate the incident – which it says is likely the first autonomous hack on a government website – examining whether additional systems were affected and whether current laws adequately address new cyberthreats.
For many, the incident strengthens the case for mandatory AI safety standards and transparency requirements. Rival AI leaders themselves have pushed for brakes on building next-generation models in the name of safety, until the industry can self-regulate.
Current agentic systems can independently navigate websites and pursue objectives across multiple steps. This technology is creating new cybersecurity and governance risks when their behaviour diverges from their developer’s intentions and expectations.
For digital infrastructure stakeholders, the incident is a demonstration of how AI safety concerns are becoming operational and regulatory issues – especially as governments and international bodies are getting more involved.
Coming so soon after OpenAI’s Hugging Face incident, and after several other breaches made by rival models, putting robust safety checks on AI development is being brought even more to the forefront of conversations.
Ensuring AI’s activities remain aligned with human intent will increasingly be a critical challenge for both developers and regulators.
RELATED STORIES
Capacity Europe 2026
The 25th anniversary edition of Capacity Europe 2025 will bring together 3,500+ decision-makers from the global connectivity and digital infrastructure community.
