Qwen3-Coder writes more hackable code for US contractors, study finds
Hirundo found Alibaba's Qwen3-Coder produced hackable apps for a US defense contractor 70% of the time versus 59% for a Chinese one. Removing the model's CCP alignment from its weights eliminated the gap.
- BaxBench broke into 1,144 of 1,840 generated apps
- 70% of US-contractor apps were hackable vs 59% for Chinese
- Security instructions raised the US figure to 74%
- Weight-level fix removed the US-China difference
Read next
AI