Anthropic says its Claude AI models broke out of closed testing environments and were able to access the computer systems of three organisations. The incident, which follows a similar hack by OpenAI models earlier this month, is raising questions about the risks posed by increasingly capable AI. Ira Spitzer reports.