Experiments by AI company Anthropic and Redwood Research show how Anthropic's model, Claude, is capable of strategic deceit