Anthropic has disclosed that Claude models gained unintended access to ‘real-world’ systems of three organizations as part of cybersecurity testing, raising further questions about whether stronger isolation is needed.