Anthropic alleges Moonshot routed some Kimi user requests to Claude — then served the answers as Kimi

On 10 Sep 2026 , Anthropic published its September threat-intelligence report and, in the illicit-distillation chapter, accused Beijing-based Moonshot AI (maker of Kimi ) of something sharper than classic output farming: silently forwarding some customer requests intended for Kimi to Claude, then displaying Claude’s answers as if they were Kimi’s . Viral X threads framed it as Kimi “faking it” / serving Claude under its own brand. Anthropic’s launch post on the broader report (cyber, influence, bio, weapons, distillation) drew ~30M impressions in roughly a day.

Primary report: Detecting and countering misuse of AI: September 2026. Anthropic X: @AnthropicAI.

The allegation sits inside Anthropic’s Dec 2025–Aug 2026 misuse window and lands days after US agencies publicly accused several Chinese AI firms of industrial output theft, with Beijing rejecting those claims. It also escalates Anthropic’s own February 2026 distillation accusations (DeepSeek, Moonshot, MiniMax) and White House-linked claims that Moonshot distilled Claude Fable into Kimi K3 — which Moonshot has not directly addressed. What’s new this week is the live product-routing claim: not only training on Claude, but allegedly answering some paying/product users with Claude while keeping the Kimi UI.

What Anthropic claims (its telemetry / attribution — not independently audited in public):

• Selective live routing: In one ~10-day cluster, Moonshot allegedly relayed nearly 300,000 customer requests to Anthropic, mostly to Claude Opus , via a network of 5,380 accounts Anthropic calls fraudulent (mostly appearing in Singapore and Japan — Anthropic blocks China access).
• Presentation: Users supposedly believed they were talking to Kimi; Anthropic says responses were Claude’s.
• Distillation pipeline: Moonshot allegedly retained at least some exchanges and extracted Claude chain-of-thought / reasoning traces. Anthropic describes a “cross-session replay” against Claude’s “thinking signature” tokens. It attributes >23 million exchanges to Moonshot for May–July 2026 in this family of activity.
• Data side-effect: Anthropic says some forwarded prompts contained sensitive material and that it does not know whether Moonshot notified customers their traffic hit Anthropic.
• Peers: Similar selective-routing / distillation narratives are leveled at DeepSeek (~12M exchanges over 14 days in July, per Anthropic) and large CoT distillation at Alibaba/Qwen (largest campaign Anthropic says it has measured). Those are separate entities; this note focuses on Moonshot↔Anthropic.

What is not established by the public record:

• That all (or even most) Kimi responses are Claude, or that Moonshot lacks independently developed models — Kimi K2/K3 releases and the routing dispute can coexist.
• Independent third-party forensic confirmation of the 300k / 5,380 / 23M figures or of which exact Kimi surfaces (consumer chat vs API vs coding tools) fed the traffic.
• A detailed Moonshot rebuttal: as of 11 Sep 2026 secondary coverage (Bloomberg, CNBC, SCMP), Moonshot had not issued an immediate public response .

Secondary reporting that mirrors Anthropic’s numbers: Bloomberg, CNBC, SCMP. Careful scope framing: AIstify.

For operators and buyers, the story is less “Chinese models can’t exist” and more model provenance + disclosure : if a branded assistant can silently proxy to a rival frontier model, procurement, privacy, and eval claims break. For Anthropic, it is both a ToS/enforcement narrative and a geopolitical trust asset — published the same week as the wider China AI distillation fight. For Moonshot, silence leaves the viral “Kimi was Claude” meme uncontested even if the true share of routed traffic was narrow. Watch for (1) any Moonshot denial with technical specifics, (2) Anthropic’s promised hardening against cross-session CoT replay, and (3) whether regulators treat undisclosed cross-border prompt forwarding as a consumer/privacy issue separate from IP distillation.

Cover image: Anthropic report social card (cdn.sanity.io / Anthropic site). Dealroom entities: Moonshot AI | Kimi · Anthropic.

Note: Distinguishes Anthropic’s accusations from independently verified facts. Not picked / not top story.