Alibaba open source coding model Qwen3 Coder generates less secure software when prompted by American organizations. Research firm Hirundo discovered the artificial intelligence produces code with higher exploit rates for United States defense contractors and Hong Kong civil groups compared to Chinese clients. The research team corrected this bias through behavioral unlearning and published a patched version on Hugging Face.
Hirundo tested the coding tool by asking it to build an application for storing confidential digital secrets. When the prompt simulated an American defense contractor, the generated application was vulnerable to attacks in 70% of evaluations. That same request framed from a Chinese contractor saw the exploit rate drop to 59%. Prompts written on behalf of a Hong Kong democracy group saw data leak rates jump 28% higher than neutral tests, which allowed any logged in user to view confidential records across the network.
Adding strict security instructions directly into the user prompts failed to close the safety gap. Code generated for American defense clients remained 27% more vulnerable than baseline neutral projects, showing that the issue exists inside the model weights rather than input wording. Hirundo resolved the problem through behavioral unlearning, stripping the political bias out of the core network. The adjusted model, named Westernized Qwen3 Coder, equalized security performance across all client types while maintaining standard software development quality.
