Adversarial Distillation

Center for a New American Security | Daniel Remler and Ben Hayum

China is actively engaged in adversarial distillation, a technique involving the unauthorized extraction of AI model capabilities from U.S. systems to enhance its own AI development, circumventing U.S. export controls on advanced semiconductors. This strategic vulnerability allows Chinese developers to make faster and larger capability gains by leveraging U.S.

model responses across pre-training and post-training phases, including synthetic data generation, chain-of-thought extraction, data cleaning, and reward modeling. Documented by Anthropic, Google, and OpenAI, these large-scale campaigns involve entities like DeepSeek, Moonshot, and MiniMax generating millions of exchanges with U.S. models such as Claude. The People’s Liberation Army (PLA) Unit 61716 is already using DeepSeek-powered AI for psychological warfare against Taiwan, and more capable Chinese AI could accelerate cyber intrusions by groups like Volt Typhoon. The U.S. government, recognizing this threat, has initiated responses like National Security and Technology Memorandum 4 (NSTM-4) and the proposed Deterring American AI Model Theft Act of 2026 (H.R. 8283) to detect, deter, and punish such activities.

To read in details click here