Moonshot AI accused of secretly using Claude

What Moonshot AI is accused of

Anthropic alleges that Moonshot AI secretly sent some Kimi user prompts to Claude, then displayed the resulting answers as output from its own system. This looks like more than ordinary model routing: it raises both product-substitution concerns for users and the possibility of extracting a competitor’s capabilities.

According to published details of the incident, nearly 300,000 requests were transferred over ten days through 5,380 fake accounts. Many were directed to Claude Opus models. Anthropic also links those exchanges to training Moonshot systems, although this remains an allegation by the company rather than an independently established violation.

Technically, such a setup is not unusual. An application receives a message, an internal router selects an external API, forwards the context and returns the answer through the original assistant interface. Moonshot’s compatible API documentation shows requests in the Anthropic Messages API format, so changing the underlying model may be as simple as replacing an endpoint and token.

A compatible interface alone does not prove abuse. It can be useful for moving tools between providers. The line is crossed when an external call is hidden from the user, access limits are bypassed, or answers are systematically collected to improve a competing model.

In a February 2026 report, Anthropic had already accused Moonshot AI, DeepSeek and MiniMax of industrial-scale distillation campaigns involving many fake accounts and millions of exchanges with Claude. This episode makes the picture more troubling: the model may have been not only a source of training data, but an invisible backend for another product.

Why hidden routing changes the equation

The main consequence concerns the origin of an answer. If the Kimi interface is effectively served by Claude without disclosure, users do not know where their prompts are sent, which data-processing rules apply or whose restrictions determine the system’s behavior.

From an engineering perspective, I would first examine recurring prompt patterns, clusters of accounts with synchronized activity, proxies, and an unusual concentration of requests involving reasoning and coding features. But there is an uncomfortable trade-off: aggressive protection against model extraction can easily block legitimate agent systems, large-scale evaluations and compatible clients.

For providers, this supports stricter account verification, anomaly analysis and answer-provenance tracking. For users, it is a reminder that the assistant name on screen no longer guarantees which model is operating underneath. The key unresolved issue is not whether platforms can build routers, but whether they can reliably prove which model actually answered a user.

We previously covered Anthropic’s reversal of silent Claude query downgrades after concerns over transparency. The Moonshot AI allegation extends that same trust question to where Claude users’ queries may actually be processed.