U.S. security institutions are alleging that half a dozen Chinese artificial intelligence organizations, featuring DeepSeek, Alibaba, alongside Moonshot AI, have utilized leading U.S. models on an industrial scale to assist in training their own technologies.
The NSA, FBI, and CISA announced on Tuesday that these firms have harvested billions of tokens across millions of interactions from models such as Claude, GPT, Gemini, and Grok since at least late 2024, frequently leveraging proxy providers and numerous profiles to sidestep identification.
The authorities characterized this behavior as malicious knowledge distillation, noting that it likely transpired with the knowledge of Beijing.
Distillation represents a standard technique for training artificial intelligence, where developers pose a high volume of queries to a superior model, gather the responses, and apply those outputs to enhance a separate model.
American officials argue that these Chinese businesses escalated this approach substantially, transmitting millions of meticulously crafted inquiries to U.S. models in order to replicate their programming, mathematical, and analytical proficiencies while operating various accounts and proxies to bypass detection and operational caps.
Moonshot serves as the creator behind Kimi K3, the open-weight system that surpassed Anthropic and OpenAI on a prominent programming benchmark during July, temporarily provoking a widespread sell-off in artificial intelligence assets.
Chinese artificial intelligence equities experienced the most severe losses, with Z.ai dropping roughly 27% and MiniMax sliding around 16%, while semiconductor stocks across the Asian region softened alongside decreases in bitcoin and other leading digital assets. The downward momentum persisted into the subsequent trading period, as South Korea’s Kospi index declined by an additional 3.5%.
CoinDesk noted during that period that Kimi’s results disputed two major premises regarding artificial intelligence. The initial premise suggested that constructing elite models would remain exceedingly costly, whereas the second assumed that superior architectures would continually originate from a limited cohort of heavily capitalized American corporations.
Moonshot demonstrated that a Chinese research lab could introduce an open-weight system capable of competing with, and occasionally outperforming, creations from OpenAI and Anthropic independent of the massive capital expenditures associated with U.S. artificial intelligence pioneers.
That realization compelled market participants to question the longevity of the American advantage, and whether the hundreds of billions of dollars dedicated to processors and server farms would persistently yield superior models.
Moonshot AI did not reply to an inquiry from CoinDesk concerning these claims during the morning hours in Asia.
Originally published at https://www.coindesk.com/tech/2026/09/09/moonshot-s-kimi-rattled-markets-u-s-agencies-now-say-it-was-trained-on-american-models.
