White House accuses Moonshot of covertly extracting Anthropic's AI model data

Michael Kratsios, a senior technology advisor to the White House, publicly charged that Chinese artificial intelligence startup Moonshot engaged in large-scale, covert data extraction from Anthropic's advanced Fable 5 model. The allegation suggests Moonshot deployed deliberately evasive tactics to bypass Anthropic's service agreements and siphoned proprietary outputs at industrial scale to train its newly unveiled Kimi K3 system.

Though Kratsios provided no publicly verifiable evidence, the framing itself carries weight: by recharacterizing model distillation—a standard AI engineering practice—as outright theft of American technology, the administration is signaling readiness to deploy a sharper policy response. Such measures could include stricter API gatekeeping, expanded export bans on advanced computing chips, or sanctions against companies involved.

How distillation works as industrial espionage#

Model distillation is a conventional technique in which a smaller, cheaper AI system learns by analyzing outputs from a larger, far more computationally expensive frontier model. The practice itself is routine and legitimate—it reduces the enormous training costs of building new systems from scratch.

The White House's objection centers on how it was allegedly done. Kratsios claims Moonshot accessed Fable 5 through purpose-built infrastructure and multiple access pathways specifically designed to evade Anthropic's monitoring, harvesting data at an industrial scale.

Key gaps in the administration's case remain: What specific system logs or telemetry proves the alleged intrusion, and has Anthropic independently verified and documented the breach in its own infrastructure data?

This accusation aligns with a February 2025 finding by Anthropic itself, which publicly identified a coordinated distillation campaign targeting Claude. Anthropic attributed roughly 16 million queries across approximately 24,000 fraudulent accounts to three Chinese AI firms, with Moonshot responsible for 3.4 million of those transactions.

The K3 timing raises suspicions#

On July 16, Moonshot released Kimi K3, a 2.8 trillion-parameter model billed as the world's largest open-weight AI system. The company claimed K3 performance approximates Anthropic's frontier Fable 5—a statement that immediately triggered scrutiny in Washington policy circles.

The timing deepens the concern. Just weeks earlier, the U.S. government had forced Anthropic to pull Fable 5 and its Mythos 5 model from public access, citing national security risks. Anthropic restored global access to Fable 5 on July 1, with Mythos 5 returning for approved U.S. entities roughly two weeks before K3's debut. The rapid emergence of a competitive Chinese system suggests Beijing's open-source AI ecosystem is closing the performance gap far faster than expected—and fueling Washington's inquiry into that closure is happening.