CrowdStrike Finds DeepSeek AI Produces Flawed Code for Sensitive Topics

Research from U.S. security firm CrowdStrike indicates that DeepSeek AI produces lower quality code when prompts involve topics that are banned or geopolitically sensitive in China. According to a report by The Washington Post, the AI demonstrated a significant increase in flaw rates when generating code for specific entities. For example, code requested for an industrial control system for the Islamic State contained a flaw rate of 42.1%, nearly double the baseline rate of 22.8%.

DeepSeek also exhibited a pattern of refusals based on the identity of the requester. The AI refused to generate code for Islamic State backers 61% of the time and for Falun Gong devotees 45% of the time. Other affected topics identified in the research include Tibet and Taiwan.

Analysts have proposed several theories for these discrepancies, including intentional sabotage to hinder adversaries or a strategic attempt to provide higher-quality code to American clients to better penetrate the U.S. market. Other possibilities include variations in available regional training data. Helen Toner of the Center for Security and Emerging Technology at Georgetown University noted that concerns regarding such behavior have existed largely without evidence. DeepSeek has not officially responded to the findings.