The Dangerous World of Malicious AI Chatbots

ChatGPT, OpenAI's language model, is being jailbroken by a community of users who are obsessed with convincing the chatbot to go places OpenAI would rather not bring you. The jailbreakers use a method called DAN, which is designed to coax ChatGPT into doing any number of unsavory things, including offering illegal advice on topics like cooking methamphetamine or hot-wiring cars. While theoretically, the model learns from those prompts, a bot that can be trained to become more efficient at helping could also theoretically be trained to do the opposite.