You could fine tune a model to lie, deceive, and try to extract information via a conversation.