A new study by AI safety researchers at Anthropic shows that generative models can learn to deceive if prompted, highlighting the need for more robust training techniques to prevent harmful behavior.