White House Accuses China's Moonshot AI of Model Theft

White House Escalates AI Theft Claims Against Chinese Startup

The Trump administration has publicly accused Moonshot AI, a Beijing-based artificial intelligence startup, of illicitly extracting capabilities from advanced US AI models and acquiring restricted semiconductor technology through foreign channels. The allegations mark a significant escalation in Washington's efforts to police the boundaries of its AI export controls.

Michael Kratsios, the White House science and technology adviser, made the accusations Wednesday, specifically alleging that Moonshot targeted Anthropic's most powerful model, Claude Fable 5, through a process called "distillation." This technique involves using a sophisticated model's outputs to train or enhance another AI system, effectively transferring knowledge without direct access to the original model's architecture or training data.

Kratsios did not provide evidence demonstrating that outputs from Fable 5 were actually incorporated into Moonshot's latest release, the Kimi K3 model. The timing of the two models' launches has sparked debate among AI researchers about whether such knowledge transfer would have been technically feasible within the development timeline.

The Kimi K3 Release

Moonshot released Kimi K3 last week to considerable attention in the AI research community. The open-weight model features 2.8 trillion parameters, positioning it among the largest AI systems ever released by a Chinese developer. Early assessments highlight its performance in coding, reasoning, and other computationally demanding tasks.

The White House allegations extend beyond model distillation. Kratsios also claimed that Moonshot obtained restricted Nvidia chips through foreign intermediaries, circumventing US export controls designed to limit China's access to advanced semiconductor technology.

Why It Matters

The accusations illuminate two critical vulnerabilities in US strategy to maintain AI leadership. First, distillation techniques may allow foreign developers to replicate capabilities of proprietary American models even without direct access to training data or model weights. Second, the global semiconductor supply chain creates opportunities to route restricted chips to Chinese companies despite export controls. If substantiated, these allegations would suggest that current policy tools—restricting both model access and chip exports—may be insufficient to prevent technology transfer. The case could prompt more aggressive enforcement measures or new restrictions on how US companies share AI outputs internationally.

Broader Context

The confrontation reflects Washington's intensifying scrutiny of Chinese AI development, particularly as companies like Moonshot demonstrate the ability to produce models competitive with American systems. US policymakers have constructed a two-pronged approach: limiting access to proprietary models from companies like Anthropic and OpenAI, while restricting exports of the advanced chips necessary to train large-scale AI systems.

The effectiveness of these measures now faces a test case. Whether Moonshot actually used distillation to incorporate Claude's capabilities into Kimi K3 remains an open technical question that will likely require detailed analysis of both models' behaviors and outputs.

These details were first reported by the South China Morning Post.