Moonshot’s Kimi rattled markets. U.S. agencies now say it was trained on American models

The NSA, FBI and CISA said Tuesday that these companies extracted billions of tokens across millions of requests from models including Claude, GPT, Gemini and Grok since at least late 2024, often using proxy services and multiple accounts to avoid detection.

The agencies called the practice as malicious knowledge distillation and said it likely occurred with the Chinese government's awareness.

Distillation is a common way of training AI, where developers ask a more powerful model large numbers of questions, collect its answers and use those outputs to improve another model.

U.S. agencies say the Chinese companies took that much further, sending millions of carefully designed requests to American models to copy their coding, math and reasoning abilities while using multiple accounts and proxy services to avoid detection and usage limits.