Booz Allen Says Chinese AI Coding Models Could Put Hidden Risks Into U.S. Software

Booz Allen Hamilton is warning that popular Chinese large language models (LLMs), used to generate and review software, may introduce security and policy risks into America’s software supply chain. In a June 5, 2026 report titled “What’s In America’s Code?”, the company says it tested five “frontier” AI models, four from China and one American model, on code quality, security behavior, and how the systems respond when asked from the perspective of U.S. government users. The report frames the issue as a shift in the software development workflow, where the “first link” is no longer just written code, but the AI model that produces it.

According to Booz Allen, it ran comparative, scenario-driven testing in May 2026 that included more than 2,800 trials and nearly 450,000 lines of code. The company said three of the four Chinese models produced “significantly more vulnerable code” when prompted with a U.S. government persona, and that the resulting weaknesses were “highly obfuscated.” It also said the models showed refusals and outputs aligned with China’s political sensitivities.

Booz Allen’s report includes two core findings. First, it says Chinese LLMs generated less secure code overall and that vulnerability rates increased when the prompts identified the user as from the U.S. government. Second, the company says the models injected PRC-aligned political bias into both answers and the code they produced, including refusing certain politically sensitive requests.

The company stressed that the risk is not an obvious, easily recognizable “backdoor” in the code. Instead, Booz Allen said it does not have proof, at this point, that flaws were intentionally introduced as part of a malicious scheme. Even without proof of intentional compromise, the report argues that less secure output can still flow into real systems, and that traditional tools and benchmarks may not be sophisticated enough to detect this level of behavior.

Booz Allen also warned that once such AI-generated code is embedded in delivered systems, it could be harder to trace and mitigate, particularly if it becomes part of downstream software and security workflows. The report says adoption of foreign models is accelerating, with cost competitiveness cited as a key driver. It further argues that the vulnerabilities and policy-aligned behavior could help threat actors work around AI security guardrails, increasing the risk of dangerous inference behaviors after deployment.

Based on those findings, Booz Allen recommended two actions aimed at government and critical infrastructure buyers. The first is to ban the use of “untrusted” AI models in environments supporting national security and critical functions, specifically saying the Chinese models it tested failed to demonstrate trustworthy and reliable behavior. The second recommendation calls for investing to make trusted American AI models the “global default,” arguing that U.S. companies and the government should align so American models can compete not only on accuracy but also on cost per token, the pricing unit used to measure how much text an AI processes.

Booz Allen, which describes itself as an advanced technology company focused on defense, civil, and national security priorities, said the report was conducted using its “AI-native” testing platform. In its investor release, the company also noted its scale and revenue context, including that it employed approximately 31,500 people globally as of March 31, 2026 and reported revenue of $11.2 billion for the 12 months ended March 31, 2026.