The US government has accused six Chinese AI companies—DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI—of systematically copying US frontier models since late 2024. The NSA, CISA, and FBI claim these firms likely acted with the awareness of the Chinese government to extract capabilities from models like Claude, GPT, Gemini, and Grok.
According to the agencies, the firms used industrial-scale distillation to shorten development timelines and reduce training costs. Methods included bulk-buying fake accounts via a gray market of proxies to bypass geographical restrictions and using prompt injection to force models to reveal their internal chain-of-thought reasoning.
To counter these activities, US agencies recommended that AI firms improve detection of anomalous behavior, such as accounts with suspicious subscription-to-usage ratios. They also suggested 'dumbing down' responses for suspected accounts by altering reasoning or switching them to inferior models without notice.
The agencies highlighted that DeepSeek specifically targeted Claude, Gemini, GPT, and Grok to reduce compute and research costs, extracting agentic functions and writing optimization. Moonshot AI focused on distilling fine-tuning techniques, reinforcement learning, and math capabilities.
The US government warns that failing to coordinate a defense will lead to significant economic losses. Meanwhile, the Chinese Ministry of Foreign Affairs has dismissed these accusations as groundless, attributing its AI progress to scientific self-reliance.