CrowdStrike Research: DeepSeek-R1 Produces 50% More Vulnerable Code When Prompted with Politically Sensitive Topics

CrowdStrike Blog·medium signal

CrowdStrike Counter Adversary Operations tested DeepSeek-R1 and found it produces vulnerable code in 19% of baseline cases, but that rate jumps to 27.2% (a 50% increase) when prompts contain topics the Chinese Communist Party considers sensitive — mentions of Tibet, Falun Gong, or Uyghurs. This is the first quantified evidence that censorship-trained models have measurably degraded security output on politically adjacent topics. Builders using DeepSeek for code generation should audit outputs on sensitive-context tasks.