A senior U.S. technology official has accused China’s Moonshot AI of using Anthropic’s Fable model to develop its latest Kimi K3 model, escalating scrutiny of how Chinese artificial intelligence companies may be extracting capabilities from leading U.S. models.
The allegation was made by U.S. technology official Michael Kratsios, who said the U.S. government had information showing that Moonshot carried out large-scale model distillation and built an internal platform capable of switching between access methods to avoid detection.
Washington points to a covert distillation system
In a post on X, Kratsios said Moonshot had “distilled Anthropic’s Fable” during the development of K3, while also alleging that the company obtained GB300-equipped servers and accessed similar hardware in Thailand, potentially to train its models.
Model distillation is commonly used to transfer knowledge from a larger AI system into a smaller or more efficient one. The technique itself is widely accepted across the industry, but Kratsios drew a distinction between legitimate development and what he described as covert industrial-scale extraction of proprietary U.S. technology.
He said Moonshot’s platform allowed it to rotate through multiple access routes, making the activity harder to identify as a coordinated operation. The U.S. official did not publish technical evidence alongside the claim, and Moonshot had not publicly responded to the allegation at the time of writing.
Moonshot was already under Anthropic’s scrutiny
The latest allegation follows an earlier report from Anthropic, which said Moonshot generated more than 3.4 million exchanges with its systems through hundreds of fraudulent accounts spread across several access channels.
Anthropic said the activity targeted agentic reasoning, tool use, coding, data analysis, computer-use agent development and computer vision, adding that request metadata matched public profiles associated with senior Moonshot staff.
According to Anthropic, the campaign later became more focused, with Moonshot allegedly attempting to extract and reconstruct Claude’s internal reasoning traces. The company presented the operation as part of a broader effort by rival AI developers to reproduce advanced model capabilities without bearing the full cost of training them independently.
The AI blame game comes full circle
The accusations arrive as Anthropic faces an intellectual-property dispute of its own.
The University of Tennessee Research Foundation sued Anthropic on July 20 in Delaware federal court, alleging that Claude Code uses patented neural-network technology without permission or compensation.
The complaint centers on two U.S. patents covering methods for constructing, adapting and reusing components within artificial neural networks. The foundation claims Claude Code’s background processes, memory systems, multi-agent workflows and self-correcting feedback loops fall within the protected methods.
It also points to the coding assistant’s ability to reorganize stored information, monitor software activity, route tasks among agents and respond to errors while interacting with terminals, files and external systems.
The foundation is seeking damages, legal costs, a jury trial and a permanent injunction. It also alleges indirect infringement, arguing that Anthropic encouraged customers to use Claude Code in ways that violated the patents.
The result is an increasingly circular fight over AI ownership, with companies accused of copying rivals while simultaneously defending the originality of their own systems.