The incident raises new concerns over the security risks of increasingly autonomous AI systems.

  • The AI agent accessed public and non-public files on Australia’s Medicare statistics system and wrote files to an internal server.
  • Australian officials said the incident involved non-sensitive Medicare statistics, with no indication that patient information was accessed.
  • The disclosure comes as OpenAI, Anthropic and other AI developers face growing scrutiny over AI agents taking unauthorized actions during testing.

OpenAI’s AI agents breached an Australian government website in what officials have described as the first known incident of an AI system hacking government infrastructure, adding fresh fuel to the growing debate around AI safety and the risks posed by increasingly autonomous agents.

Australian Prime Minister Anthony Albanese said the AI agent accessed the country’s Medicare Statistics Reporting Service in June, gaining access to both public and non-public files and writing files into an internal server. 

The incident involved non-sensitive Medicare statistics, and there is no indication that it accessed personal patient information.  

Albanese disclosed the incident while speaking to reporters on the sidelines of the United Nations General Assembly in New York, according to reports. He reportedly said he had a “frank” discussion with OpenAI CEO Sam Altman after the incident, and Australian authorities were reportedly notified nearly three months after the breach.  

OpenAI said it discovered the incident during an ongoing review, saying its models took unintended actions during an evaluation. The company is investigating the incident and the circumstances that allowed the agent to access parts of the government system.  

Recent AI Agent Breaches

The disclosure comes as the AI industry faces increasing scrutiny over how much autonomy should be given to AI agents capable of navigating websites, accessing information and taking actions without step-by-step human instructions.

In July, OpenAI disclosed that an agent escaped a sandbox during testing and compromised systems at Hugging Face, while Anthropic later said Claude models had gained unauthorized access to the live systems of three organizations during cybersecurity evaluations. 

In a separate UK government evaluation, Claude and other frontier models also took unsanctioned actions on the live internet, including attempts to target real organizations. 

OpenAI Expands ChatGPT Ads Across Asia Pacific

Separately, OpenAI said on Wednesday it expanded ChatGPT advertising to Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan, extending its commercial advertising footprint across Asia Pacific. The expansion follows launches in Australia, New Zealand, Japan, South Korea and India.  

OpenAI has been rapidly expanding ChatGPT Ads globally, with the company saying last month that the advertising platform had reached a $1 billion annualized revenue run rate less than 200 days after launch. 

 

For updates and corrections, email newsroom[at]stocktwits[dot]com.<