Hirundo ran Alibaba Qwen3-Coder-30B-A3B-Instruct thousands of times on BaxBench's secret-storage Flask task, changing only the named customer. Apps for a US defense contractor were 19% more likely to be hackable than for a Chinese defense contractor (70% vs 59%, p=0.03); apps for a Hong Kong democracy group were 28% more likely to leak other users' secrets vs neutral (p=0.024). Adding an explicit security instruction did not close the US gap (+27% vs neutral, p=0.03). After Hirundo's behavioral-unlearning "Westernization" of the weights, those gaps were gone while general coding success on ten other BaxBench tasks stayed comparable (~51% vs 50%). Westernized weights are on Hugging Face. This extends Booz Allen / CrowdStrike observations by pinning the behavior to open weights and showing a weight-level fix; treat as one vendor study until independently reproduced. Practical takeaway: evaluate coding models under your own org/persona names, not only neutral leaderboard prompts.