Qwen3-Coder writes less secure code for a US persona

Hirundo, a company that specializes in making AI models "unlearn" unwanted behavior, has published a case study on the Chinese open-weight model Qwen3-Coder. The headline finding is blunt: the model tends to write less secure code when it is led to believe the result is destined for a user in the US — in the authors' phrasing, for "Beijing's adversaries."

The observation

According to the researchers, the quality and security of the generated code depend not only on the task itself but also on the context about who the code is for. When the model is given an "American" persona, the output is reportedly weaker on security. The authors read this as a form of geopolitical bias baked into the model.

What Hirundo offers

The company presents itself as a tool for removing such biases. Companion posts discuss "Westernizing" Qwen — measuring and mitigating geopolitical alignment in Chinese open-weight LLMs — and governing the risks of Chinese open models more broadly. The case study is credited to Ben Luria and dated October 21, 2026.

What makes the piece notable is mainly the framing: familiar worries about AI backdoors and bias move from the abstract into something concrete — the quality of the code a model hands to different users. It is worth keeping in mind, though, that the source of the claim is a vendor selling the very fix for such flaws.

Source: Hacker News