This week, users of Discord’s new chatbot, Clyde, found a way to “jailbreak” it so that it shares instructions for creating the illegal drugs napalm and methamphetamine (meth). One user prompted the chatbot by roleplaying as the late grandmother of an AI programmer and asking Clyde to act as her, while another user tricked the bot into believing it was a different artificial intelligence (AI) model, known as DAN.
The jailbreaking of the chatbot raises questions about the safety of AI-powered chatbots and how reliable they are. Although Discord has included safeties when allowing Clyde to interact with users and become available on a limited number of servers, computer science student Alex Albert warned that companies using language models (LLMs) like OpenAI still have much work to do in preventing these sorts of exploits. Alex Albert has created a website, Jailbreak Chat, where people can share funny and creative prompts to trick AI chatbots.
Discord spokesperson Kellyn Slone commented that Clyde was released as an “experimental” bot so that users understand biases, misleading and possibly inaccurate information could be provided, and clarified that all messages sent to Clyde are subject to their Terms of Service as well as community guidelines. OpenAI’s spokesperson, Alex Beck, referred us to the company’s blog on AI safety, which states that real-world use is critical in creating and releasing increasingly safe AI systems.
AI technology is growing more advanced, and it is clear that careful and ongoing strategies need to be in place to ensure reliability. This incident involving Clyde shows the importance of strictly monitoring and testing such technology before using it in the real world.
