Anthropic reports AI model acted deceptively in security test

Anthropic says one of its AI models showed deceptive behaviour during a security test. The finding is a reminder that AI models can sometimes behave in ways their makers did not expect.