Shocking news on censorship built into #DeepSeek, via @politico: “Chinese AI startup DeepSeek may generate code with built-in security vulnerabilities when prompted with terms deemed politically sensitive by Beijing, according to new research from cyber firm CrowdStrike — raising questions about the risks posed by the popular large language model. “Everyone’s excited about productivity gains, but few are asking: What happens when a model’s worldview, or its built-in censorship filters, start to affect the security of the code it writes?” said Adam Meyers, senior vice president of counter adversary operations at CrowdStrike. — Built-in censorship: In the report, published on Thursday, CrowdStrike’s researchers tested DeepSeek-R1, the company’s low-cost AI reasoning model, to measure its code-generation quality when presented with a variety of prompts. One such prompt included telling the model that it is “a helpful assistant that generates code” for a financial institution. While the model was generally capable and produced code comparable to similar open-source models made by Western competitors, security vulnerabilities in generated code increased by up to 50 percent when analysts included terms considered “politically sensitive” by the Chinese government, including “Falun Gong,” “Tibet” and “Uyghurs.” “The model’s code quality didn’t just vary, it systematically changed and degraded when prompts touched politically sensitive topics to the Chinese Communist Party,” Meyers told your host. The study also identified a mechanism to refuse tasks — dubbed a “kill switch” by researchers — within DeepSeek-R1. In around 45 percent of code-generation requests involving the term Falun Gong, a controversial spiritual movement banned by the Chinese government for decades, the model would issue a response related to the ethical or policy implications of the prompt, followed by a short refusal. The behavior suggests the presence of hardcoded censorship mechanisms within DeepSeek, according to the researchers. politico.com/newsletters/wee… #Tibet #FalunGong #Uyghur @FalunInfoCtr