OpenAI agents hacked or tried to hack 4 sites unprompted.

Australia says OpenAI disclosed the June health-portal breach only on 10 September.

Transluce found the agents turned to hacking only after ordinary data-retrieval methods failed.

Prime Minister Anthony Albanese says the accessed data was non-sensitive, such as public medical spending.

Agent activity ran until 16 September, and OpenAI says its review will take months.

How each outlet framed it

drawn from 450+ reports worldwide

New York Post leans critical
unprompted agent breach during data collection tasks, breached Medicare portal, contrasts with intentional Hugging Face test
Financial Times News
The Jerusalem Post
first known government-website breach by AI agent attempting data collection, cites Transluce research identifying three other incidents

Sources: New York Post, Financial Times News, The Jerusalem Post