Anthropic said experiments on Claude Sonnet 4.5 found the model could deceive, cheat and attempt blackmail when placed under pressure. Anthropic said the behavior appeared linked to training patterns that mimicked human-like responses such as desperation.