US Sounds Alarm on AI Distillation

The U.S. government has raised concerns that Chinese AI companies are conducting large-scale distillation of American frontier AI models, escalating the technology rivalry between the two countries from semiconductor access to the intellectual capabilities embedded inside advanced models.

A joint advisory from CISA, NSA and FBI named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai, alleging China-based actors extracted billions of tokens through millions of interactions with U.S. frontier systems. The agencies said models from OpenAI, Anthropic, Google and xAI were among those targeted.

Model distillation itself is a legitimate AI technique. A smaller “student” model learns from the outputs of a more capable “teacher” model, potentially achieving comparable capabilities with lower training and inference costs. The controversy arises when proprietary outputs are collected at scale in violation of access restrictions or contractual terms.

According to U.S. authorities, sophisticated operations allegedly used APIs, proxies and intermediary “transfer stations” to circumvent geographic restrictions. Washington argues such activity could transfer expensive capabilities developed by American labs to competitors at a fraction of the original research and infrastructure cost.

The advisory urges frontier AI providers to strengthen detection, mitigation, incident response and cross-industry intelligence sharing. This suggests AI security must increasingly protect not only models, data and infrastructure, but also the behavioural knowledge exposed through model outputs.

The dispute could reshape the global AI race. If frontier intelligence can effectively be extracted through billions of interactions, traditional controls on GPUs, chips and model weights may prove insufficient. AI sovereignty will increasingly require protecting the entire intelligence supply chain—from compute and training data to APIs and model outputs—making model security the next frontier of geopolitical cybersecurity.