{"slug": "anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude", "title": "Anthropic Says Six Other Chinese AI Labs Did What Xiaomi Did to Claude", "summary": "Anthropic's Case GTG-16008 threat report names seven China-based AI labs — Xiaomi, Alibaba, DeepSeek, Moonshot, Zhipu, SenseTime and MiniMax — as running industrial-scale campaigns to extract Claude outputs for their own training pipelines, generating roughly 190 million exchanges with Claude. Anthropic says Xiaomi alone routed more than 400,000 requests through 1,500 accounts over a 20-day window in March and April 2026, and that Moonshot forwarded close to 300,000 requests from its own Kimi users to Claude and passed the answers back without telling them. Xiaomi shipped MiMo-V2.6 on Hugging Face on September 22, scoring 46 on Artificial Analysis's Intelligence Index, first among the 114 open-weight models tracked, while China's Cyberspace Administration has opened a formal probe into DeepSeek and Moonshot over routing more than 35 million combined user requests to Claude.", "body_md": "*Xiaomi wasn't the only one. Anthropic's full threat report names Alibaba, DeepSeek, Moonshot, Zhipu, SenseTime and MiniMax alongside Xiaomi, and Moonshot's Kimi was caught quietly routing its own users' questions through Claude without telling them.*\n\nWhen Startup Fortune covered Xiaomi's MiMo-V2.6 topping the open-weight leaderboard last week, the story looked like a two-party dispute: one Chinese phone maker, one American AI lab, one accusation of stolen homework. Anthropic's fuller threat report, published this month and detailing what it calls Case GTG-16008, makes clear it's a lot bigger than that. Seven China-based labs, not one, ran what Anthropic describes as industrial-scale campaigns to extract Claude's outputs and feed them into their own training pipelines. Together they generated roughly 190 million exchanges with Claude.\n\nXiaomi's slice of that, according to Anthropic, involved more than 400,000 requests. They were routed through 1,500 accounts over a 20-day window in March and April 2026, funneled through proxy services designed to mask where the traffic was really coming from. Anthropic says the pattern points to a specific playbook. Xiaomi allegedly launched a free trial for international developers, let usage build, then used that access to reconstruct developer environments from transcripts, clean up multi-turn conversations into structured training pairs, and generate synthetic exchanges that mimicked real human-model conversations. The bulk of the distillation, Anthropic says, happened right as the trial period was ending.\n\nXiaomi has not responded to any of it. Not to the original accusation, not to the fuller report, and not to the fact that it shipped MiMo-V2.6 on Hugging Face on September 22, less than two weeks after Anthropic went public with the claim. The model's flagship Pro version scored 46 on Artificial Analysis's Intelligence Index, putting it first among the 114 open-weight models the firm tracks, tied with Grok 4.7 and 20 points above its own predecessor. Silence, in a story this specific, reads as its own kind of answer.\n\nThe detail that got less attention than it deserved is what Anthropic calls model rerouting, and it's a different kind of violation from distillation. According to research cited in Anthropic's report, Moonshot forwarded close to 300,000 requests from its own Kimi users to Claude and passed the answers back. Those users believed they were talking to Moonshot's own model the whole time. DeepSeek and Xiaomi ran comparable relays. Distillation is a company training on another company's outputs. Rerouting is a company selling its own product while quietly outsourcing the actual thinking to a competitor, without telling the customer who's really answering.\n\n[China Probes DeepSeek and Moonshot Over Secret Data Routing to Claude](https://startupfortune.com/china-probes-deepseek-and-moonshot-over-secret-data-routing-to-claude/)\n\nAnthropic's September 10 threat report accused DeepSeek and Moonshot AI of secretly routing over 35 million combined user requests to Claude models, some containing Chinese police and military-linked data. China's Cyberspace Administration has now opened a formal probe into both companies, just days before planned Trump-Xi AI talks. - [China AI companies secretly routing data to Claude](https://startupfortune.com/china-probes-deepseek-and-moonshot-over-secret-data-routing-to-claude/) - [how DeepSeek and Moonshot funneled requests to Claude](https://startupfortune.com/china-probes-deepseek-and-moonshot-over-secret-data-routing-to-claude/)\n\nThat distinction matters for anyone evaluating an open-weight Chinese model on paper. A benchmark score tells you nothing about whether the answers came from the model you downloaded or from a live pipe to somebody else's model during testing. Kyle Chan, a researcher who has tracked the report closely, pointed out on X that Anthropic's own data cuts against a simple narrative. Zhipu tried to distill Anthropic's Fable model and failed, reportedly because Fable's safeguards held, and switched to distilling Claude Opus 4.6 instead specifically because its defenses were weaker. That's not a story about Chinese labs being incapable of building their own models. It's a story about which targets are easiest to copy.\n\nAnthropic's response has been to harden Fable 5.1 with a feature it calls preserved thinking. It cryptographically locks a conversation's context, including the system prompt and available tools, so a new API session can't tamper with what preceded Claude's reasoning. The model now also summarizes its internal reasoning before responding, which makes a stolen transcript less useful as training data even if someone captures it whole. It's a defensive arms race playing out inside the product itself, not just in press statements.\n\nNone of Anthropic's claims have been tested by a court or a regulator, and that caveat still stands. But the accusation is no longer Anthropic's word against one company's silence: it's Anthropic's word against seven companies' silence. There's a documented pattern across labs that includes at least one case, Zhipu's, where the company's own behavior lines up with what Anthropic says happened - try one target, fail, switch to an easier one. For founders and investors treating open leaderboards as a shortcut to picking a model, the lesson isn't that Chinese labs can't build good models. Several clearly can. It's that a leaderboard position tells you nothing about how the model got there. And right now the honest answer, according to the company whose product allegedly got copied, is that for at least seven labs it didn't get there alone.\n\n**Also read:** [Lovable Says Its Annualized Revenue Just Crossed $600 Million](https://startupfortune.com/lovable-says-its-annualized-revenue-just-crossed-600-million/) • [Meta's Six-Week Muse Blitz Reclaims Its Spot At AI's Frontier](https://startupfortune.com/metas-six-week-muse-blitz-reclaims-its-spot-at-ais-frontier/) • [Meta Unveils Muse Charm, a Pocket AI Companion With No Price Yet](https://startupfortune.com/meta-unveils-muse-charm-a-pocket-ai-companion-with-no-price-yet/)\n\n*This article is posted in [AI News](https://startupfortune.com/category/ai/), check it out for more related stories.*\n\n## Join the discussion\n\n[Open in the community →](https://startupfortune.com/community/)\n\nAlmost there. Sign in and your reply posts straight away.\n\n[Xiaomi's MiMo-V2.6 Tops Open AI Charts Amid Anthropic Distillation Claims](https://startupfortune.com/xiaomis-mimo-v26-tops-open-ai-charts-amid-anthropic-distillation-claims/)\n\nXiaomi's new MiMo-V2.6 model just topped the global open-weights AI leaderboard, built for roughly $2.62 million. The release came less than two weeks after Anthropic accused Xiaomi of routing over 400,000 conversations through Claude to help train it, one of seven Chinese labs named in Anthropic's latest threat report. - [Xiaomi MiMo-V2.6 open weights model benchmark](https://startupfortune.com/xiaomis-mimo-v26-tops-open-ai-charts-amid-anthropic-distillation-claims/) - [Anthropic Claude distillation campaign allegations Xiaomi](https://startupfortune.com/xiaomis-mimo-v26-tops-open-ai-charts-amid-anthropic-distillation-claims/)", "url": "https://wpnews.pro/news/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude", "canonical_source": "https://startupfortune.com/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude/", "published_at": "2026-09-24 17:41:50+00:00", "updated_at": "2026-09-24 18:00:46.574176+00:00", "lang": "en", "topics": ["ai-safety", "large-language-models", "ai-policy", "artificial-intelligence", "ai-ethics"], "entities": ["Anthropic", "Xiaomi", "Alibaba", "DeepSeek", "Moonshot", "Zhipu", "SenseTime", "MiniMax"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude", "markdown": "https://wpnews.pro/news/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude.md", "text": "https://wpnews.pro/news/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude.txt", "jsonld": "https://wpnews.pro/news/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude.jsonld"}}