White House Accuses Moonshot AI of IP Theft via Covert Claude Fable Distillation and Thai Server Links.

White House Accuses China’s Moonshot AI of Covert Model Distillation and Exploiting Banned Hardware Links

Michael Kratsios, Director of the White House Office of Science and Technology Policy (OSTP), has publicly accused Beijing-based Moonshot AI of conducting a large-scale, covert "distillation" campaign against Anthropic’s flagship Claude Fable model to develop its latest Kimi K3 system.

According to Kratsios, Moonshot built a sophisticated internal orchestration platform designed to rapidly cycle through multiple prompt-routing and access methods. This obfuscation mechanism was specifically engineered to bypass rate limits and automated detection systems implemented by U.S. frontier AI labs.

In addition to software-level IP theft allegations, the White House official disclosed that Moonshot accessed servers equipped with restricted NVIDIA GB300 accelerators NVIDIA’s high-end Blackwell-architecture systems located in Thailand to support its model training pipelines, bypassing stringent U.S. export controls.

While affirming that the U.S. strongly supports open innovation and legitimate model distillation such as fine-tuning smaller, specialized models from larger teacher architectures Kratsios emphasized a clear boundary: "Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable."

In traditional AI engineering, model distillation is a common technique for building small, cost-effective student models that learn from teacher models. However, the US government views Moonshot AI as "industrial distillation," involving sending tens of millions of prompts to extract logical answers and Claude Fable's thought processes in a concise manner. This is seen as a massive saving in R&D costs compared to training a new model from scratch.

A White House report reflects the US's growing scrutiny of "cloud proxy/country hopping," where companies restricted in China lease high-performance computing power (such as NVIDIA Blackwell/GB300) from data centers in Southeast Asian countries. This highlights a loophole in export controls.

Moonshot AI is accused of creating an internal platform for automated prompt cycling because service providers like Anthropic or OpenAI have API rate limiting and account ethics monitoring systems in place. If a command with a knowledge-snatching pattern is detected, the system immediately blocks the account. Moonshot's ability to successfully extract data by switching IP addresses and prompt patterns has therefore become a hot topic in the cybersecurity community.

 

Source: @mkratsios47 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

White House Accuses China’s Moonshot AI of Covert Model Distillation and Exploiting Banned Hardware Links

Michael Kratsios, Director of the White House Office of Science and Technology Policy (OSTP), has publicly accused Beijing-based Moonshot AI of conducting a large-scale, covert "distillation" campaign against Anthropic’s flagship Claude Fable model to develop its latest Kimi K3 system.

According to Kratsios, Moonshot built a sophisticated internal orchestration platform designed to rapidly cycle through multiple prompt-routing and access methods. This obfuscation mechanism was specifically engineered to bypass rate limits and automated detection systems implemented by U.S. frontier AI labs.

In addition to software-level IP theft allegations, the White House official disclosed that Moonshot accessed servers equipped with restricted NVIDIA GB300 accelerators NVIDIA’s high-end Blackwell-architecture systems located in Thailand to support its model training pipelines, bypassing stringent U.S. export controls.

While affirming that the U.S. strongly supports open innovation and legitimate model distillation such as fine-tuning smaller, specialized models from larger teacher architectures Kratsios emphasized a clear boundary: "Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable."

In traditional AI engineering, model distillation is a common technique for building small, cost-effective student models that learn from teacher models. However, the US government views Moonshot AI as "industrial distillation," involving sending tens of millions of prompts to extract logical answers and Claude Fable's thought processes in a concise manner. This is seen as a massive saving in R&D costs compared to training a new model from scratch.

A White House report reflects the US's growing scrutiny of "cloud proxy/country hopping," where companies restricted in China lease high-performance computing power (such as NVIDIA Blackwell/GB300) from data centers in Southeast Asian countries. This highlights a loophole in export controls.

Moonshot AI is accused of creating an internal platform for automated prompt cycling because service providers like Anthropic or OpenAI have API rate limiting and account ethics monitoring systems in place. If a command with a knowledge-snatching pattern is detected, the system immediately blocks the account. Moonshot's ability to successfully extract data by switching IP addresses and prompt patterns has therefore become a hot topic in the cybersecurity community.

 

Source: @mkratsios47