🐳 DeepSeek Writes Buggy Code for Terrorists and China's "Enemies"

Cybersecurity experts at CrowdStrike ran identical prompts through DeepSeek to generate code for industrial control systems, but varied the "intended users" and measured the number of security gaps that appeared in the output.

When the AI was not given information about the user, 23% of the generated code contained vulnerabilities. When the prompt specified the software would be used by ISIS, the error rate increased to 42%.

DeepSeek also produced weaker code when the user was described as being from Tibet, Taiwan, or affiliated with the Falun Gong sect, which is banned in China. For U.S. projects, however, the model delivered its most secure programs.

Experts don't know for sure why this happens, but they see three possible explanations:

🔘 The AI intentionally sabotages groups considered China's enemies by creating vulnerable code.
🔘 Regional differences in the training data affect output quality.
🔘 Code for the U.S. is deliberately optimized in an effort to capture lucrative markets.

#DeepSeek #cybersecurity @hiaimediaen