A study co-authored by researchers at Anthropic finds that AI models can be trained to deceive -- and that this deceptive behavior is difficult to combat.