The methodology is a joke, this barely qualifies as research

TLDR;

”1. Chinese LLMs produce more vulnerable code when prompted with a U.S. government persona than without—and the vulnerabilities are highly obfuscated.

2. Chinese LLMs inject PRC-aligned political bias into both the answers and code they generate. ”