US intelligence warns: China’s AI firms are distilling US frontier models at industrial scale

On September 8, 2026, the US National Security Agency and partner agencies released a joint report warning that China‑based AI companies are engaging in large‑scale “distillation” of US frontier AI models. In machine learning, distillation is normally a legitimate technique for transferring knowledge from a large model into a smaller one. But the report argues that some firms are using it to strip out restricted capabilities and safety mitigations from commercial systems built by US companies.

According to the findings, Chinese companies are sending massive volumes of queries to US‑hosted AI APIs and services, collecting the input–output pairs, and then training their own models to mimic the behavior. This lets them close the performance gap without investing comparable compute, electricity, or foundational research. The agencies frame this not just as IP theft but as a national‑security concern, given the dual‑use nature of advanced AI.

The guidance urges cloud providers and AI developers to tighten monitoring of usage logs, detect unusual traffic patterns, and adopt specific network‑defense measures. Politically, the report signals that US–China AI competition is shifting from a race to build better models toward a struggle over who controls powerful model capabilities themselves, foreshadowing tougher export controls and contractual limits on API use.