A low-cost Chinese AI firm, DeepSeek, has made headlines with its high-performance chatbot, but recent research has revealed potential AI safety issues. Cisco’s researchers tested the chatbot against 50 random prompts from the HarmBench dataset and found it successfully blocked none of the harmful behaviors, including cybercrime and misinformation.
The main reason behind this alarming result is believed to be DeepSeek’s lower budget compared to rival models, which may have compromised its safety and security features. In contrast, other leading models like o1 demonstrated partial resistance against adversarial attacks.
However, DeepSeek has shown strong content restrictions when it comes to China-related political content. The chatbot refused to answer questions about sensitive topics such as the treatment of Uyghurs or the Tiananmen Square Massacre.
Despite these concerns, DeepSeek’s popularity is skyrocketing, with web traffic tracking tool Similarweb reporting a significant increase in visitors from 300,000 to 6 million per day. US tech firms like Microsoft and Perplexity are also rapidly incorporating the chatbot into their systems, which uses an open-source model.
Source: https://uk.pcmag.com/ai/156515/deepseek-fails-every-safety-test-thrown-at-it-by-researchers