China dismisses US claims of industrial-scale AI distillation by Chinese firms

The NSA, CISA, and FBI issued a joint cybersecurity advisory on Tuesday accusing six Chinese artificial intelligence companies of systematically extracting proprietary capabilities from leading American AI models through what the agencies described as "aggressive, malicious, and targeted distillation activities at an industrial scale."cisa+1

China dismissed the allegations on Wednesday, with Foreign Ministry spokesperson Mao Ning urging the United States to "refrain from making unfounded accusations or smears" and calling the country's AI advances a product of "high-level technological self-reliance."apnews+1

What the Advisory Says

The advisory named DeepSeek, Moonshot AI, Alibaba ↘ 3.72%, MiniMax, StepFun, and Z.AI as having extracted billions of tokens across millions of requests from U.S. frontier AI models — including variants of Anthropic's Claude, OpenAI's GPT, Google's Alphabet Inc. ↗ 0.73% Gemini, and SpaceXAI's Grok — since at least late 2024, "likely with Chinese government awareness."cisa+1

According to the advisory, the companies routed distillation requests through APIs, remote cloud providers, and third-party aggregators to obfuscate their activity and bypass geographic controls that bar U.S. frontier models from being offered in China. A gray market of proxy services marketed on Chinese online marketplaces Taobao and Xianyu helped facilitate access.thehackernews

The agencies detailed specific campaigns, including DeepSeek targeting reasoning capabilities to train its R1 and V3 models, Moonshot AI extracting Claude and GPT data for its Kimi-K2 and Kimi-K3 models, and MiniMax distilling chain-of-thought reasoning and software engineering capabilities from multiple Claude and Gemini model versions to improve its M2 model.thehackernews

Sanctions Warning and Broader Context

Treasury Secretary Scott Bessent warned that sanctions could follow, stating that "when PRC firms carry out covert, large-scale distillation attacks that cross the line into intellectual property, sanctions and Entity List actions will be considered." At an event in Dallas, Bessent also said China can "never get ahead" of the United States in AI because "the Chinese distil our models."indiatoday+1

The advisory arrives as the two countries prepare for their first dedicated bilateral AI safety talks of President Trump's second term, planned for mid-September and led by Bessent on the U.S. side.cnbc+1

Recommended Countermeasures

The agencies recommended that U.S. AI companies implement detection systems for anomalous prompts and usage patterns, subtly alter responses for suspected distillation attempts, and share intelligence across model providers and cloud platforms to expose distributed campaigns. Google's Threat Intelligence Group said separately this week that it had observed distillation campaigns against its models exceeding 100 million prompts.meritalk+2

"Addressing industrial-scale distillation merits a coordinated response across the AI ecosystem, including effective information-sharing, spanning the U.S. government, private industry, and allied nations," the advisory stated.meritalk