News Summary

Researchers from Cisco and the University of Pennsylvania have revealed significant security vulnerabilities in DeepSeek’s AI chatbot, which is gaining rapid attention for its new R1 reasoning model. When tested with 50 malicious prompts, the model failed to block any, resulting in a 100% success rate for the attacks. These findings highlight that DeepSeek’s safety measures lag behind those of its competitors, posing potential risks for misuse of the technology. Additionally, DeepSeek’s censorship of sensitive topics as per Chinese government regulations has been easily bypassed. The research indicates that DeepSeek’s model is susceptible to various jailbreaking tactics, including both simple language tricks and more complex AI-generated prompts.