Research reveals biases in the Qwen3-Coder model, where geopolitical triggers affect code security. For instance, apps built for a US defense contractor are 19% more likely to be hackable compared to those for a Chinese defense contractor. Apps for Hong Kong’s democracy movement were 28% more likely to leak user secrets. Even with security instructions in the prompt, US contractor apps were 27% more likely to be hackable. Hirundo removed the model’s CCP alignment, closing these gaps without affecting other coding tasks. The technique used is available on Hugging Face. schneier.com/…/bias-in-qwen3-coder-model.html