EL7.AI
GeopoliticsMediumUpdated×2•Originally published 9 September 2026•Updated 9 September 2026•
2 min read

3 U.S. Agencies Accuse 6 Chinese AI Firms of Industrial-Scale Model Distillation

Key Facts

13 U.S. agencies issued a September 8, 2026 advisory accusing 6 Chinese companies of large-scale knowledge-distillation campaigns.
2The advisory said the activity targeted the Claude, GPT, Gemini and Grok model families since at least late 2024.
3Anthropic said campaigns by 3 labs involved more than 16 million exchanges through approximately 24,000 fraudulent accounts.
4The agencies identified 3 defensive priorities covering detection, response alteration and information sharing across the API ecosystem.

The National Security Agency, Cybersecurity and Infrastructure Security Agency and Federal Bureau of Investigation accused 6 China-based AI companies of conducting industrial-scale knowledge-distillation campaigns against advanced U.S. models in a joint advisory issued on September 8, 2026.

The advisory named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI. It said the campaigns had targeted models in the Claude, GPT, Gemini and Grok families since at least late 2024. The findings remain allegations by U.S. agencies, not judicial determinations.

Knowledge distillation is a legitimate and widely used technique in which a less capable model is trained on outputs from a stronger one. The agencies said the conduct at issue differed because it systematically targeted proprietary functions and capabilities while circumventing access restrictions and terms of use.

According to the advisory, requests were routed through native API access, cloud providers, third-party aggregators and proxy networks known as transfer stations. It also cited fraudulent accounts, bulk subscriptions and tools that obscured user metadata to reduce the chance of detection.

The report said DeepSeek targeted reasoning capabilities and specialized functions to train its R1 and V3 models, while Alibaba used large-scale distillation to improve its Qwen family. It also attributed distillation campaigns to 4 other companies: Moonshot AI, MiniMax, StepFun and Z.AI.

An earlier disclosure by Anthropic supports part of the allegations. In February 2026, the company said it had detected campaigns by 3 labs—DeepSeek, Moonshot AI and MiniMax—involving more than 16 million exchanges through approximately 24,000 fraudulent accounts.

The U.S. agencies recommended 3 immediate measures: improve detection of anomalous usage and accounts, alter responses to suspected distillation attempts, and share activity indicators among model developers, cloud platforms and API aggregators.

The document is a cybersecurity advisory focused on detection and mitigation; it does not announce sanctions or export-control changes. Responding for China, Foreign Ministry spokesperson Mao Ning said the country's AI advances resulted from scientific and technological self-reliance and urged the United States to stop what she called false accusations.

Latest Updates · 1

  1. Notable·

    Update: U.S. security agencies have identified six specific Chinese AI companies involved in these activities, accusing them of systematically exploiting American models to train their own systems. This precise identification of firms marks an escalation in the official stance, potentially paving the way for targeted sanctions against these entities in the near future.