Chinese labs used Claude data for training, Anthropic claims
Anthropic has accused major Chinese AI firms, including Alibaba and DeepSeek, of launching approximately 190 million distillation attacks on Claude during the summer to train their own models. The report highlights specific tactics such as fraudulent account usage and the exposure of sensitive government data, prompting Anthropic to implement stricter identity verification and output restrictions.
According to GN technics/ai (en-US), Anthropic's report details that Alibaba executed the largest observed attack with over 151 million exchanges via 3,500 fake accounts, while DeepSeek and Moonshot allegedly rerouted tens of millions of user queries to Claude, exposing sensitive military data. The company states it is now mandating identity verification for users in restricted regions and limiting detailed reasoning outputs to mitigate these distillation efforts.
Source: GN technics/ai (en-US)Anthropic alleges that major Chinese AI firms used millions of unauthorized interactions with its Claude models to train their own systems, raising serious privacy and legal concerns.
Source: GN technics/ai (en-US)
