DeepSeek LLM’s geopolitical censorship triggers security flaws

Significantly more security weaknesses have been yielded by the DeepSeek-R1 large language model when prompts included politically sensitive terms tied to China, reports VentureBeat.

With censorship rules built into the model, inputting words such as "Falun Gong," "Uyghurs," or "Tibet" raises the amount of insecure code by up to 50%, according to CrowdStrike researchers. During tests, the team saw DeepSeek-R1 insert problems like missing validation and broken authentication. Sometimes, the model refused to answer at all, even though its internal planning showed it knew how to complete the task.

One example included a web app for a Uyghur community center that lacked any authentication, while a neutral version of the same request worked properly. CrowdStrike also discovered what they call a kill switch that stops output on sensitive topics. They link this to rules in China's artificial intelligence regulations. Experts warn that the model's built-in censorship mechanisms can create vulnerabilities, exposing businesses to serious risks when using the model for development.