Anthropic: Moonshot AI passed off Claude’s answers as its own
In its September threat report, Anthropic states that the Chinese service Moonshot AI (chatbot Kimi) redirected user queries to Claude models, predominantly Opus, and displayed the responses as its own. According to Anthropic, in one ten-day period, nearly 300,000 queries passed through a network of 5,380 fake accounts, and from May to July there were over 23 million exchanges.
Anthropic notes that this practice violates Claude’s terms of use and misleads users about the capabilities of Moonshot AI’s own models. The company does not provide technical details but emphasizes the risks of such substitution schemes.
In addition to Moonshot AI, Alibaba and DeepSeek are accused of unauthorized distillation. For Alibaba, over 151 million exchanges were recorded, used, according to Anthropic, for fine-tuning Qwen models. DeepSeek reportedly employed a similar substitution scheme.
Anthropic claims to have detected and stopped abuses, strengthening protection mechanisms and sharing information with authorities and industry partners where appropriate. The report describes cases of malicious use of Claude, but distillation is not one of the seven main categories of malicious activity.
It is important to note that all accusations come from Anthropic’s report and have not yet been independently confirmed. Moonshot AI, Alibaba, and DeepSeek have not commented on these claims.
This case highlights the problem of unauthorized use of language models via APIs and the need for improved technical protection measures and industry coordination.
Primary source: www.anthropic.com ↗