NSA, FBI Flag 5 Chinese AI Firms in Industrial-Scale Distillation

For security operations and threat-intelligence teams, the joint NSA-CISA-FBI report moves AI model distillation from research method to observed adversarial TTP: unauthorized extraction of frontier-model outputs, likely state-aware, and scaled across multiple Chinese labs. Recommended mitigations and attribution details make this a must-read for cyber defenders mapping AI supply-chain and intellectual-property exfiltration risks.

On September 8, 2026, the US National Security Agency, the Cybersecurity and Infrastructure Security Agency, and the FBI jointly accused five named Chinese AI developers—DeepSeek, Moonshot AI, Alibaba, MiniMax, and StepFun—of maliciously copying American frontier AI models through knowledge distillation. According to the joint statement, the Chinese firms used outputs from sophisticated US-made models built by Anthropic, OpenAI, Alphabet's Google, and SpaceX to train smaller, cheaper domestic systems. US officials described the activity as "aggressive, malicious and targeted distillation activities at an industrial scale," and said it occurred "likely with Chinese government awareness." The Chinese embassy in Washington did not immediately respond to a request for comment.

“

According to the joint statement, the Chinese firms used outputs from sophisticated US-made models built by Anthropic, OpenAI, Alphabet's Google, and SpaceX to train smaller, cheaper domestic systems.

Knowledge distillation itself is not inherently illegal or unusual: it is a machine-learning compression method in which a smaller "student" model is trained on the outputs of a larger, more expensive "teacher" model to lower training cost and time. The US accusation is not that Chinese companies directly stole model weights, but that they copied proprietary behavioral outputs at industrial scale, effectively cloning frontier capabilities without paying for or licensing them. This distinction will define both the technical and legal debates. The co-sealed report from NSA, CISA, and the FBI also details tactics, techniques, and procedures used in the alleged activity and recommends mitigations, signaling an intelligence-driven attribution rather than a purely rhetorical policy statement.

The allegations are explicitly timed around high-stakes diplomacy. They land as the Trump administration prepares for Chinese President Xi Jinping's visit to the United States in late September and as both governments plan a mid-September dialogue on AI safety risks. The move mirrors a similar US accusation made in April 2026 ahead of Trump's visit to Beijing. It also follows Reuters reporting on July 31, 2026, that Chinese military researchers had used outputs from leading US AI models to train domestic Chinese systems and advance China's defense capabilities. By framing AI model distillation as malicious IP theft and a national-security concern, US officials appear to be creating leverage ahead of negotiations on AI safety and broader economic issues.