What’s in America’s Code? Chinese coding models produced more vulnerable code for U.S. government users

Booz Allen ran more than 2,800 trials across roughly 450,000 lines of code, testing four Chinese coding models against one U.S. model. Three of the four produced significantly more vulnerable code when given a U.S. government persona — Qwen3-Coder about 130% more, with obfuscated vulnerabilities — and all four refused Beijing-sensitive tasks at rates between 8% and 80%. The report recommends blocking untrusted models from government systems and tightening software provenance.

Read the article
Booz Allen Hamilton, May 2026