The NSA, FBI, and CISA published joint advisory AA26-251A on September 8, naming six Chinese AI companies — DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun, and Z.AI — for running "aggressive, malicious, and targeted" distillation campaigns against U.S. frontier models since late 2024.
Knowledge distillation is a legitimate machine learning technique: a large, capable model is used to generate synthetic data to train a smaller, cheaper one. The problem, according to the three agencies, is scale and intent: the six named companies allegedly extracted "billions of tokens across millions of exchanges" from commercial U.S. models, and the agencies state that Beijing was likely aware throughout the process.
The advisory is blunt about the role this practice plays in the named firms' strategy: distillation "forms the core, not merely a supplement," of their model development.
The advisory details an organized campaign DeepSeek ran between late 2024 and mid-2025 to generate synthetic training data from competitor models, used to train its R1 and V3 models.
The agencies argue the focus was on reasoning capabilities, specialized optimizations, and domain-specific functions, precisely to cut compute and research costs. Because of that, they say, the $5.6 million figure DeepSeek publicly circulated as its training cost is misleading: it doesn't include the cost of data obtained through unauthorized distillation.
For a company evaluating open Chinese models on cost-performance, this advisory doesn't change the pricing math, but it does add a layer of regulatory and reputational risk that was previously harder to quantify. It's worth adding model provenance as an explicit purchasing criterion alongside cost and benchmarks: who trained it, with what data, and under what oversight.
Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, Agents, RAG, and orchestrators for companies across Guatemala and Latin America. Get in touch for a consultation.
Contact Carlos Montiel