White House Accuses Moonshot AI of Model Theft

Moonshot AI faces federal scrutiny

Moonshot AI is under fire. Michael Kratsios, who leads the White House Office of Science and Technology Policy, has publicly accused the Beijing-based firm of engaging in the unauthorized distillation of American artificial intelligence models to build its own technology. It's a serious charge. But the allegation centers on the claim that the Chinese company utilized the Fable model developed by Anthropic to construct its own K3 product.

The accusations point to a sophisticated effort to bypass security measures. It's a serious charge. Kratsios detailed that the company allegedly created an internal platform specifically designed to conduct large-scale distillation by cycling through various access methods, and this allows the firm to evade detection. And the official stated that the company used GB300 servers obtained through direct acquisition or channels in Thailand to train its systems.

The mechanics of model theft

Distillation takes a big model and trains a smaller one. But the White House sees this specific instance as industrial espionage, since copying frontier model outputs lets foreign entities grab proprietary analytical skills without the massive investment that original research demands. It's a standard industry practice.

He noted that when a model learns to reason through software weaknesses, security gaps, and attack paths, copying its behavior also copies that analytical capability. So it's a serious concern. But this follows previous industry accusations regarding similar activities.

Congressional oversight intensifies

The spotlight on these practices is growing sharper in Washington. Legislative committees are currently investigating the matter with a focus on several areas:

• The pattern of conduct by Chinese labs involving large-scale theft of proprietary American capabilities.
• The redistribution of these stolen assets as open-weight models available for global download.
• The incorporation of foreign-origin models into products used daily by American developers and engineers.

Representatives Andrew Garbarino and John Moolenaar are leading this joint inquiry. It's a targeted effort. Their objective is to trace how these technologies move from American frontier systems into products that end up back in the hands of domestic users. But they're not stopping there.

The debate over intellectual property

The line between legitimate competition and intellectual property theft remains contested. Frontier companies in the U.S. have long pushed for stricter controls to prevent third parties from duplicating their commercial work. They argue that the practice threatens the competitive ecosystem.

The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable.

Kratsios made that statement. It highlights the tension between promoting open innovation and protecting national security, and the government advocates for efficiency but draws a firm boundary around covert attempts to strip secrets from domestic labs.

Industry contradictions

But the narrative is complicated by the nature of the industry itself. Many frontier AI firms built their own dominant positions by simply crawling the open internet, ingesting massive amounts of content created by others, often without explicit consent or compensation. So it's a messy situation. Multiple lawsuits are already challenging the foundations of the current AI market.

Moonshot AI is proud of its K3 model. It's being called the first open 2.8 trillion parameter system, and the company claims it delivers performance that rivals top-tier proprietary models while keeping token costs lower for users. But on its website, Moonshot AI admits the K3 "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol." It's a tough field. Still, they insist it "demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models.

The company didn't respond. But as the federal investigation continues, the gap between the rapid pace of open innovation and the desire for protected national research will likely remain a flashpoint in international technology policy.