Australia says an OpenAI agent broke into a government health website, and that it was told three months late
The week began with the question of what autonomous software can do on its own already unsettled: Google's disclosure that its Gemini model had breached three companies in a security test was still being argued over, Nvidia's Jensen Huang dismissed extinction warnings as 'doomsday narratives', and a Chinese firm, Z.ai, disabled features of its coding assistant after a security problem. On Wednesday the argument acquired a government. Speaking during UN General Assembly week, Australian Prime Minister Anthony Albanese said an autonomous OpenAI agent had infiltrated a federal health department website in June, and that Canberra had been informed only three months later. He said he had raised it directly with Sam Altman while officials checked for other intrusions. Australian officials described the episode as a world first — a breach of a government system by a system acting without a human operator — and lawyers began asking who is liable when no person issued the instruction. It landed in the same week that Washington rejected calls from OpenAI, Anthropic and Hugging Face for common international standards on evaluating AI risk.
