Anthropic traces nearly 300,000 Kimi requests to an Opus model through 5,380 fraudulent accounts

Anthropic says it does not allow its technology to be accessed from inside China ^[9]. That restriction is what makes this an account story. The traffic Anthropic described to Bloomberg came through accounts that mostly appeared to operate from Singapore and Japan ^[2], and Anthropic's report described proxy operators maintaining pools of accounts and replacing them when individual accounts were blocked ^[10].

Divide the two Bloomberg figures and the pool shows up in the arithmetic. Nearly 300,000 requests across 5,380 accounts is about 56 requests per account ^[12]. That is thin per-account volume for a data collection run at this scale. From the provider's side of the API, that pattern looks like accounts being cycled and replaced.

The two published figure sets do not nest. Anthropic's own report describes more than 3.4 million exchanges through hundreds of fraudulent accounts ^[3]; the narrower cluster Bloomberg describes carries 5,380 accounts ^[2], which is under nine percent of the exchanges and more of the accounts ^[13]. The source material does not say whether that is a different counting rule or a different population.

Attribution rests on request metadata that Anthropic said matched public profiles of senior Moonshot personnel ^[5]. That signal exists only inside the model provider. A cloud reselling weights, or a developer building on Kimi, has access only to the response and the bill, not to the metadata behind it.

Moonshot has been seeking up to 30% of revenue generated from K3 services on Azure, AWS and Google Cloud ^[7], after launching the 2.8-trillion-parameter mixture-of-experts model on July 16th and releasing its weights on July 27th ^[6]. A revenue-share arrangement on open weights is a contract about a model family. The routing allegation asks a different question: which model handled the user's prompt and produced the output the product displayed ^[8].

Distillation on its own is a standard training method, including when a provider distills its own systems ^[11]. The allegation concerns false accounts, proxy infrastructure and restricted access used to collect training data from a competitor at scale ^[4]. Routing requests among models is also ordinary practice, and it depends on disclosure, valid access and agreements governing how user data is processed ^[14].

What the public record does not carry is independent confirmation that Kimi's consumer product forwarded user prompts to Claude ^[15]. The timing of the alleged traffic is also unestablished, so nothing ties it to K3 or to any other Kimi release ^[16]. And the secondhand account differs from the primary one on who else is accused: Bloomberg reported that Anthropic accused DeepSeek and Xiaomi of similar conduct, while Anthropic's report named DeepSeek and MiniMax alongside Moonshot ^[17].

So the provenance question a buyer can act on is contractual. That means specifying which model serves the request, under whose access, and with what data terms in writing ^[14]. These numbers come from Anthropic's own request logs ^[5].