Three U.S. security agencies accused six China-based artificial-intelligence companies of industrial-scale intellectual-property theft, saying they trained their models on the outputs of advanced American systems without authorisation.
In a joint statement, the National Security Agency, the Cybersecurity and Infrastructure Security Agency and the FBI said the firms are engaging in 'aggressive, malicious and targeted distillation activities at an industrial scale.' Distillation is a standard machine-learning technique in which a smaller, cheaper model is trained to imitate a larger, more capable one. The agencies argue that applied to frontier American models at this scale — and without permission — it amounts to copying that lets rivals skip years of expensive research.
The companies named include DeepSeek, Moonshot AI, Alibaba, MiniMax and StepFun, part of a group of six firms the officials say used variants of American-made models to train their own products, 'likely with Chinese government awareness.'
The timing is deliberate. Washington is preparing for Chinese President Xi Jinping's visit to the United States in late September, and the two governments are due to discuss AI safety risks in a dialogue planned for mid-September. The distillation claim hands U.S. negotiators a fresh grievance in a relationship already shaped by export controls on advanced chips and curbs on cross-border AI investment.
The statement did not identify which American models were allegedly used. U.S. labs have increasingly tried to police how their outputs are consumed, tightening API terms, adding watermarks and watching for bulk querying — measures that distillation makes harder to enforce.
Independent observers note that distillation is not itself illegal and is widely used inside U.S. labs as well. The dispute is therefore as much about access and scale as about the technique itself — and it is likely to sharpen as the two countries sit down to talk about safety.
