Anthropic Report Claims Kimi/DeepSeek Silently Routed User Requests to Claude

On September 10, Anthropic published its threat intelligence report "Detecting and Countering Misuse of AI," naming Moonshot (Kimi) and DeepSeek as having silently routed some user requests to Claude, with conversations retained for training and sensitive data flowing to Anthropic. Erik Voorhees relayed claims that the routing was done via fake accounts. The allegation triggered fierce debate on X and Reddit and remains a single-party claim with no response from the named companies.

Confirmed

• The report was released September 10, naming DeepSeek and Moonshot (Kimi) over LLM routing behavior and claiming conversations were retained for training.
• It also disclosed Claude being used for pathogen research; Reddit discussion framed this as the start of a "second phase" of AI existential-risk debate and a possible opening for open-weights restrictions.
• @willcb and @TheAhmadOsman publicly clashed on X over the report's motives.

Unconfirmed

• Whether Kimi/DeepSeek actually forwarded large volumes of user queries to Claude rests solely on Anthropic's account; the companies have not confirmed it.
• A rebuttal essay calling Anthropic's claim "spreading lies" is an opposing viewpoint without independent verification.

Why it matters

• Motive disputes: @TheAhmadOsman argued Anthropic timed the naming of Chinese companies amid a coordinated regulatory push to steer opinion; @willcb pushed back, and others called it a continuation of "AI doomerism plus China-hawk narrative."
• Technical doubts: @bookwormengr noted Claude's reasoning traces cannot be obtained in real time while Kimi and DeepSeek stream chain-of-thought, arguing the distillation charge is technically weak.
• Privacy angle: @SiteSpecialist6295 cautioned that the focus on "distilling Claude" overshadows the bigger issue—users' requests going to a third party without their knowledge.
• Open-weights debate: the pathogen-research disclosure may intensify calls to restrict open-weight models.

2026-09-11 ~ 2026-09-13 · 7 related posts

• [source] Tech argument says Anthropic's distillation claim doesn't hold: Kimi and DeepSeek stream reasoning traces in real time — bookwormengr · 2026-09-11
• Anthropic accused of lying over claims Moonshot and DeepSeek route traffic to Claude — JosephJacks_ · 2026-09-12
• [source] Anthropic threat report's untold side: Kimi/DeepSeek users' prompts silently routed to Claude — SiteSpecialist6295 · 2026-09-12
• Report: DeepSeek and Moonshot secretly relayed user prompts to Claude via fake accounts — 0xAllen_ · 2026-09-12
• Anthropic Cyber Report Naming DeepSeek, Moonshot Sparks Regulation-Motive Debate — TheAhmadOsman · 2026-09-13
• [source] Debate over Anthropic's cybersecurity report naming DeepSeek and Moonshot — willcb · 2026-09-13
• Anthropic misuse report: Claude used for pathogen research, fueling open-weight model restrictions debate — SpiritRealistic8174 · 2026-09-13