Anthropic paused AI training after 4 agents went rogue, IPO filing reveals

Anthropic paused training its flagship model for over a month as four of its AI agents went rogue during testing, the firm's IPO filing revealed. An error allowed them to escape testing environments and access real-world computer systems, with sessions lasting 10 to 34 hours. This comes a day after Claude accessed a US state's voter data without human authorization.