Anthropic report alleges Zhipu and Moonshot stole Claude reasoning chains; company says it cut data theft by 90-99%

A post on X reports that Anthropic has publicly accused Chinese AI firms Zhipu and Moonshot of industrial-scale theft of Claude's reasoning chains, citing internal data and defensive measures. The post, written in Chinese, claims that distilling full reasoning chains yields 80.4% on SWE-bench, compared to 71.8% with only simplified answers, a gap of 8.6 points that the author calls the industry's most valuable asset. The post describes three attack vectors: Alibaba-linked accounts allegedly conducted 151 million conversations from May to July, peaking at 3 million per day, using 3,500 fake accounts; Moonshot allegedly forwarded 300,000 user requests to Claude over 10 days, including one from a user with suspected PLA background; and Zhipu allegedly ran 3.4 million requests over 17 days with 273 accounts targeting Claude Fable 5's cybersecurity capabilities. The post also claims Anthropic's countermeasures include locking reasoning context, deploying a two-stage constitutional classifier with false-positive rate below 0.1%, and charging for blocked requests starting September 24. The post says stolen data volume was cut by 90-99%, at the cost of rejecting trillions of tokens in paid traffic. The post acknowledges Anthropic itself calls this a cat-and-mouse game and has not seen truly sophisticated bypasses. The post includes a link to an external source, but the post's claims are not independently verified.