Advisory AA26-251A, published September 8, 2026 by the NSA, CISA, and FBI, names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI for extracting "billions of tokens across millions of exchanges/requests" from U.S. frontier models including Claude, GPT, Gemini, and Grok since at least late 2024. The agencies said the firms likely acted with Chinese government awareness and that the practice provides "significantly shorter AI development timelines and reduced financial expenditures in training a frontier model."
Moonshot AI is named for distilling 17 U.S. models, including Anthropic's Claude Fable 5, to train Kimi-K3. DeepSeek allegedly crafted prompts that forced models to expose hidden chain-of-thought reasoning steps, transferring reasoning methodology rather than just final outputs. The advisory argues DeepSeek's publicized low training cost is misleading because it excludes training data allegedly obtained through distillation.
The advisory's recommended mitigation is for U.S. providers to secretly serve suspected distillers "less sophisticated downgraded models" and to "avoid informing" them of the switch. The detection signatures named, including sustained 24/7 usage, new accounts at maximum throughput, and cache-optimized traffic, also describe legitimate production agent fleets, leaving each provider to decide which customers receive an undisclosed downgrade. No American AI company has publicly confirmed it degrades responses to suspected distillers.
Researchers demonstrated that frontier lab APIs had implementations allowing systematic extraction of reasoning traces, and Anthropic confirmed Chinese labs used this technique. Nathan Lambert described the distillation debate as "the single most eventful debate around open models in 2026," characterizing the political response as panic not grounded in evidence.
China's Ministry of Commerce called the allegations baseless, characterized distillation as a "neutral" technique used broadly including by American companies, and warned of retaliation if Washington uses the allegations to contain China's AI development. China also accused Washington of using the claims to assert technological dominance, monopolize computing power, and suppress competition.
The White House OSTP issued Memorandum NSTM-4 on April 23, 2026, which a Cloud Security Alliance analysis describes as the first U.S. government policy instrument to formally classify systematic capability extraction from frontier AI systems as a national security threat. NSTM-4 narrowly targets covert, unauthorized extraction campaigns that violate terms of service and use infrastructure designed to evade detection, not legitimate distillation.
A technical analysis contends that Chinese models including Kimi K2.5, GLM-5, and DeepSeek V3.2 had developed advanced agentic capabilities independently by early 2026 and that "the value of distilling frontier models' agentic trajectories as a cold-start strategy had diminished significantly by this point." A Lawfare analysis argues mass distillation warrants a policy response but urges policymakers to focus on securing access, sharing threat information, and identifying fraudulent accounts and proxy networks rather than "creating new quasi-IP rights in model outputs."
Nvidia CEO Jensen Huang argued in an Axios interview that "distillation, learning from AI, learning from other people, and learning from other sources of knowledge, is fundamental to intelligence," noting that AI now generates more content than humans and predicting AI-to-AI learning will become unavoidable. Independent analyses cited by Lambert estimate the performance gap between open and closed frontier AI models has narrowed to roughly 4-6 months, with leading open models coming primarily from Chinese labs since approximately 2024. Multiple major Western companies including DoorDash, Airbnb, Anysphere/Cursor, Apple, Perplexity, and Thomson Reuters have faced scrutiny from U.S. lawmakers after adopting Chinese AI models. Anthropic resigned from the Information Technology Industry Council, citing disagreement with the trade group's opposition to AI export controls.