AIBRIEF
Distillation is a standard AI training technique. Three U.S. agencies now say six Chinese labs ran it 'without permission' at industrial scale, weeks before U.S.–China AI talks.
Three U.S. agencies — the NSA, CISA, and FBI — accused six China-based AI companies on Tuesday of running "industrial-scale distillation campaigns" against U.S. frontier models, including variants of Claude, GPT, Gemini, and Grok.
The named firms are DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI. The agencies say the alleged activity has been running "since at least late 2024," involves "billions of tokens" (units of text a model processes) "across millions of exchanges," and was done "likely with Chinese government awareness." They called the campaigns "aggressive, malicious, and targeted."
Distillation is a standard AI development technique: a smaller "student" model learns to mimic a larger "teacher" model, transferring capability without copying the original. The U.S. position is that doing it without the lab's permission crosses into industrial-scale IP extraction. Anthropic argues that "illicitly distilled models lack necessary safeguards, creating significant national security risks."
Industry signals have been mixed. OpenAI told the Financial Times it had evidence DeepSeek performed distillation "against its wishes." But CEO Sam Altman said in July: "I would rather people not distill from us, for sure. But this is not in my top ten list of worries."
The statement lands weeks before Treasury Secretary Scott Bessent is expected to meet Chinese officials for AI security talks this month, per Reuters reporting.