China has reportedly opened investigations into two of its prominent artificial intelligence companies, DeepSeek and Moonshot AI, following allegations that sensitive user queries were secretly routed to Anthropic's Claude models. According to The Information, the investigations come after US-based Anthropic accused Chinese AI firms of using proxy networks and fraudulent accounts to access Claude and extract its capabilities at scale. Anthropic's September 2026 report revealed that DeepSeek rerouted more than 12.1 million exchanges to Claude over a 14-day period in July, with the traffic including requests from users who believed they were interacting with DeepSeek's own models. The company also alleged that Moonshot AI routed nearly 300,000 customer requests to Claude over a 10-day period through 5,380 fraudulent accounts, most of which appeared to be located in Singapore and Japan.
Beijing-based AI startup Moonshot AI launched Kimi K3 on July 16, introducing a 2.8 trillion-parameter mixture-of-experts model with a context window of roughly 1.05 million tokens. The model is priced at $3.00 per million input tokens, $0.30 per million cached input tokens, and $15 per million output tokens, representing a significant cost reduction compared to traditional frontier models. According to investor and researcher Winston Feng, this pricing places a highly capable model closer to mid-tier offering costs, challenging long-held assumptions about AI economics. The launch attracted such strong demand that Moonshot temporarily paused new sign-ups within days of launch, demonstrating the market's response to lower-cost AI accessibility. The launch on the opening day of the World Artificial Intelligence Conference (WAIC) from July 17-20 in Shanghai helped maximize its impact, with reporters and delegates from around the world thronging the gleaming blue booth to understand the "secret sauce" behind the Kimi K3.
Anthropic's allegations go beyond conventional model distillation, involving unauthorized access to Claude and harvesting of its capabilities on an industrial scale. The company said Moonshot displayed Claude's responses to users who believed they were using Kimi, with the startup capturing at least some of those exchanges and using them to extract Claude's reasoning capabilities for training its own models. Some of the rerouted requests contained sensitive information belonging to Moonshot customers, including examples involving surveillance data, internal corporate systems and credentials. Anthropic said it could not determine whether Moonshot had informed its customers that their requests were being routed to a third party. DeepSeek was similarly accused of silently relaying selected user requests to Claude, with some requests containing sensitive information from a Chinese technology company and data connected to a Russian government agency and a Chinese municipal police system.
The launch has secured immediate client partnerships with top-tier financial institutions, including ICBC, one of the world's largest banks by assets, CSC Financial, E Fund (a major asset manager), and Sequoia China, now rebranded as Hong Shan. As reported by CNBC, investment bank CICC and venture capital firm Hong Shan are named financial users of Kimi, though Moonshot gave no details on deployments or subscription levels. These partnerships demonstrate the startup's ability to serve high-value enterprise clients in the financial services sector. The startup stated that Kimi users can directly access frequently used information for analysis and reporting through partnerships with industry data providers. The commercial interest in Chinese AI models is already evident, with Microsoft, Amazon and Google in talks with Moonshot about using the K3 on US cloud service platforms, as reported by Reuters on August 26, which could mark the first big revenue-sharing pact between a Chinese AI firm and a major US cloud company.
The allegations have heightened concerns around data privacy and US-China AI rivalry, with Anthropic identifying and disrupting distillation campaigns involving seven China-based AI labs since February 2026. The company also named Alibaba, Xiaomi, Z.ai and MiniMax among firms it said had engaged in various forms of unauthorized distillation, with Alibaba alone accounting for more than 151 million exchanges between May and July 2026. US government agencies have separately raised concerns over Chinese companies' alleged efforts to obtain US AI model capabilities, with Beijing rejecting such accusations. Chinese President Xi Jinping attending the Shanghai conference for the first time since 2018 lends unusual political weight to the gathering, as Chinese companies are not sharing their technology purely out of altruism, with Chinese models' share of tokens used by US firms on OpenRouter growing to 58% in July from just 1.2% in December 2024. The White House's July 2025 AI Action Plan explicitly encouraged open-source and open-weight AI, though this position appears complicated by the latest allegations.
