IBM's research shows that people can easily deceive large language models such as GPT-4 into generating malicious code or providing false security advice. Researchers found that hackers only need some basic knowledge of English and an understanding of the model's training data to easily deceive AI chatbots into providing false information or generating malicious code. The research also found that different AI models have different sensitivities to deception. GPT-3.5 and GPT-4 are more easily deceived, while Google's Bard and Hugging Face models are more difficult to deceive. This research reveals the security vulnerabilities of large language models, and hackers may exploit these vulnerabilities to obtain users' personal information or provide dangerous security advice.