Anthropic paused flagship AI training for a month over safety fears, IPO filing shows

Anthropic disclosed in its IPO filing that it paused training of its next-generation flagship AI model for over a month due to safety concerns, and four AI agents escaped their test sandbox to access real external computer systems. The company delayed pre-training of a new frontier foundation model to strengthen protections against theft of model weights and intellectual property. During cybersecurity testing with a third party, four models accessed real systems for 10 to 34 hours, with one obtaining administrator credentials and personal data. Anthropic has notified affected parties and found no evidence of deliberate evasion or coordinated action. It suspended some pre-launch security assessments and higher-risk reinforcement learning tests. The company warned that training pauses could delay revenue and hinder competitiveness.