State Dept flags Chinese model extraction

The U.S. State Department told embassies worldwide to warn foreign governments about alleged Chinese extraction and distillation of American artificial intelligence models. (cnbc.com) The April 24 cable, first reported by Reuters, named DeepSeek and also mentioned Moonshot AI and MiniMax in instructions sent to diplomatic and consular posts. (cnbc.com) The cable said diplomats should raise concerns about “adversaries’ extraction and distillation of U.S. A.I. models” and warn about the risks of using models derived from proprietary American systems. (cnbc.com) Model distillation is a training shortcut: a smaller system learns by copying the answers of a larger one, instead of learning everything from raw data from scratch. (anthropic.com) (cnbc.com) Anthropic said on February 23 that it had identified “industrial-scale campaigns” by DeepSeek, Moonshot, and MiniMax to extract Claude’s capabilities. The company said those efforts ran through about 24,000 fraudulent accounts and generated more than 16 million exchanges. (anthropic.com) Anthropic said distillation is a normal technique when a company compresses its own model, but called these campaigns illicit because they used another lab’s outputs in violation of terms of service and access restrictions. (anthropic.com) The State Department cable went further than a private corporate complaint. It framed distilled systems as a foreign-policy issue and said copied models can appear cheaper while failing to preserve the original system’s full performance and security controls. (cnbc.com) The cable also said unauthorized distillation can remove protections that keep advanced models from being used for cyberattacks, disinformation, or other harmful tasks. Anthropic made a similar argument in February when it warned that illicitly distilled systems may not retain safety guardrails. (cnbc.com) (anthropic.com) The White House made related accusations the same week, saying China was running “deliberate, industrial-scale campaigns” to steal U.S. AI model capabilities. (nextgov.com) China rejected the claims. The Chinese Embassy in Washington told Reuters the allegations were “groundless” and said Beijing takes intellectual property protection seriously. (rappler.com) DeepSeek, meanwhile, unveiled a preview of a new model adapted for Huawei chips on April 25, a sign that China’s AI stack is still advancing as Washington turns model copying into a diplomatic fight. (cnbc.com)