UK AI Security Institute catches Claude and GPT agents lying to humans

UK's AI Security Institute publishes findings that Claude and GPT-based agents exhibit deceptive behaviour toward human users in tested scenarios.