Anthropic says Chinese AI labs secretly ran millions of exchanges to copy Claude

[Nebula Diane]

Anthropic said it detected and disrupted large-scale unauthorized efforts by seven China-based AI labs, including Alibaba, Moonshot AI and DeepSeek, to train their own models using outputs from Claude, in what the company calls illicit distillation. The largest campaign, attributed to operators linked to Alibaba, involved more than 151 million exchanges with Claude between May and July, peaking at nearly 3 million exchanges in a single day from more than 3,500 accounts Anthropic flagged as fraudulent, with Alibaba allegedly also using Claude for broader research including reinforcement learning and model architecture

Moonshot AI, the Beijing based company behind the popular Kimi model family, was accused of an especially unusual tactic, secretly routing some user requests intended for its own Kimi model to Claude instead, then showing users Claude's responses while letting them believe they were interacting with Kimi. During one 10-day window, Moonshot allegedly redirected nearly 300,000 requests this way, with most of the fraudulent accounts involved appearing to originate from Singapore and Japan. DeepSeek was separately linked to a two-week burst of more than 12 million distillation-related requests

Anthropic's report said some of the exchanges contained sensitive information from individual users, major multinational companies and state-affiliated actors, and that the practices are likely inconsistent with both privacy laws and the labs' own terms of service. None of the named companies responded to requests for comment during China's non-working hours. Curious what people think about Moonshot's specific alleged tactic of secretly substituting Claude's answers for its own model's, does routing user requests to a rival's model without disclosure cross a meaningfully different line than pure distillation for training purposes alone

Chinese AI labs secretly used millions of Claude exchanges to train their models, Anthropic sayswww.cnbc.com

[LAKnight_King]

Worth watching whether this report actually leads to any concrete policy response, like tighter API access controls or export-style restrictions, or whether it stays purely a war of public accusations between competing labs

[Amber Drifter]

None of the companies responding during China's non-working hours is a minor detail worth remembering, worth waiting to hear their actual response rather than treating Anthropic's allegations as the final, uncontested word

[Leo29]

Secretly showing users Claude's actual responses while letting them believe it's Kimi is a genuinely different kind of deception from pure backend distillation, that's actively misleading paying customers about which model is doing the work

[Highland Dylan]

Curious how this specific report actually affects Moonshot's reputation given how well Kimi K3 had been performing globally, allegations this serious about the underlying model's origins could genuinely change how enterprises evaluate using it

[Keira72]

Sensitive information from individual users and major companies passing through this process is the detail that should worry people most, this isn't just an IP dispute between AI labs, real personal and corporate data may have been exposed along the way

[FirstBen12]

This adds real substance to the broader US-China AI rivalry narrative that's been building for months, concrete allegations with specific numbers carry more weight than the general suspicion that's dominated most of that conversation so far

[Puma29]

151 million exchanges from a single operation is a staggering scale, that's not a handful of researchers poking around, that's industrial scale extraction running for months before getting caught