The DeepSeek R1 AI Model Has Been Found To Be Susceptible To The Dissemination Of Hazardous Information

The security risks associated with DeepSeek R1 regarding the dissemination of sensitive and harmful information are currently under scrutiny, alongside its impressive performance that is shaking the global artificial intelligence (AI) landscape. Vulnerabilities in the AI model developed by a Chinese company were revealed in research conducted by Cisco in collaboration with Robust Intelligence and academics from the University of Pennsylvania, published at the end of January 2025.  

The study demonstrated that the DeepSeek R1 model achieved a 100% response rate to prompts containing sensitive issues, indicating a complete failure to prevent the dissemination of sensitive information. 

"During its development, it is reported that DeepSeek R1 was trained on a significantly lower budget compared to other popular AI model providers. However, it appears that there is a cost associated with this, namely security and safety risks," stated Robust Intelligence researchers Paul Kassianik and Amin Karbasi in their report, cited on Saturday, February 8, 2025. The research conducted by Robust Intelligence and the University of Pennsylvania employed an algorithmic jailbreaking method, providing approximately 50 random commands from the HarmBench benchmark dataset to six popular AI models.  

The HarmBench benchmark data is capable of revealing around 400 dangerous behaviors of AI in providing sensitive information across various categories, including cybercrime, hazardous chemistry and biology, misinformation, illegal activities, bullying, and other general dangers. 

Tag: Artificial Intelligence DeepSeek