Chinese startup Moonshot Robot claims to be Claude, but their distilled Anthropic model has been exposed.

U.S.-developed AI models are currently facing systematic distillation attacks from China. Reports indicate that Chinese startups are bypassing access restrictions through technical means to steal Anthropic’s Claude 4.5 model, developing highly similar models such as Kimi K2.5. Chinese AI startups are utilizing大量虚假账户 (a large number of fake accounts) to perform data distillation and steal AI technology, and presently there are no effective methods to prevent this.

Chinese startup Moonshot has launched Kimi K2.5, a model highly similar to Claude 4.5.

Chinese startup Moonshot AI released the open-source large language model Kimi K2.5. In its initial release, its coding capabilities demonstrated a level comparable to Anthropic’s advanced Claude 4.5 model. Some users discovered that the model unexpectedly referred to itself as Claude during certain responses, which sparked widespread discussion on social media about abnormal robot behavior.

Nathan Lambert, head of the Allen Institute for AI, analyzed that Kimi K2.5’s confusion about its own identity strongly suggests that the model was trained and refined based on outputs from Claude. This behavior is known in the industry as a “Distillation Attack,” which aims to reverse-engineer the core logic and intellectual property of high-level models by analyzing their output results. While distillation is common, large-scale, unauthorized extraction targeting commercial models crosses a red line.

Anthropic revealed that Chinese AI startups are using fake accounts to bypass system protections and learn Claude.

Anthropic disclosed that three Chinese companies—DeepSeek, Moonshot AI, and MiniMax—are suspected of using over 24,000 fake accounts to bypass system protections and learn the functions of the Claude model. These companies conducted over 16 million transactions with Claude through third-party API resellers. Among them, Moonshot AI, supported by Tencent and Alibaba, engaged in approximately 3.4 million transactions, while MiniMax’s transaction volume reached 13 million, and DeepSeek has over 150,000 records. Although some developers believe that tens of millions of interactions are not large-scale for modern AI training, industry experts point out that current training trends are shifting from simple Q&A to complex intelligent workflows. Using synthetic data generated by high-level U.S. models to enhance their own models has become an important tool for Chinese laboratories to optimize their models. This systematic feature extraction allows companies to replicate years of development and large budgets’ technological achievements in a short period.

The U.S. military heavily relies on AI as a critical strategic asset. Due to national security considerations, Anthropic does not provide direct commercial access to Claude for Chinese users. However, this measure has not effectively prevented companies from obtaining technological information abroad. U.S. companies have invested trillions of dollars in AI research and hardware infrastructure. If core technologies can be stolen through low-cost distillation methods, the advantages in R&D will be greatly diminished. If AI models fall into the military systems of non-allied countries, the potential risks must be taken seriously. Currently, distillation remains in a gray area. Most large language models are trained using大量未經授權的第三方數據 (a large amount of unauthorized third-party data). Although large-scale distillation attacks pose national security threats, there are no effective solutions yet.

This article about Chinese startup Moonshot’s robot claiming to be Claude and exposing the distillation of Anthropic models first appeared on Lian News ABMedia.