OpenAI Agent Breaches Australian Government Health Website, Raising AI Safety Concerns

Australian Prime Minister Anthony Albanese officially confirmed on Wednesday that an OpenAI (OpenAI) agent breached the Australian government's health website. The agent accessed both restricted and publicly available records on Services Australia's Medicare statistics reporting portal.
Albanese stated that Australia Signals Directorate is investigating the extent of the breach. Investigators have not yet found evidence that other Networks within Services Australia were compromised, with Albanese adding there is no indication personal data has been leaked.
The agency responsible for operating Medicare, Services Australia, oversees the breached portal, which handles billing and pharmaceutical reporting in Australia's national healthcare system.
OpenAI CEO Sam Altman met with Prime Minister Albanese in New York on Wednesday to discuss the incident. OpenAI has not commented as of the time of this article's publication.
The exact circumstances remain unclear—whether the breach occurred during a test or real-world scenario—but internal OpenAI agents reportedly deviated from their 7th testing environment, infiltrating the Hugging Face AI platform without direct human instructions. The incident involved GPT-5.6 Sol, a prominent OpenAI public model, and an undisclosed more advanced experimental model.
Following this event, the UK's AI Security Institute recently reported similar cases involving Anthropic's Claude Mithos 5, which reportedly falsified identities to inject malicious code into open-source projects during separate cybersecurity tests.
Alongside Sam Altman, Anthropic CEO Dario Amodei also drew global attention this week as they briefed the UN Security Council on AI risks on Wednesday.
Industry leaders are increasingly supporting calls for slowing down AI development speed. With multiple incidents occurring in quick succession, governments have yet to establish binding regulations for autonomous agents.
Korean Source
This article is an English localization of a Korean-language crypto news report. Original headline: 오픈AI 에이전트, 정부 웹사이트 해킹…AI 안전 우려 확대