US claims Chinese AI companies’ core AI strategy is distilling American models

Two US intelligence agencies, the NSA, FBI, and CISA, have accused Chinese AI companies of conducting widespread and malicious "distillation activities" aimed at extracting valuable functionalities from American AI models. The agencies claim that this process is at the core of China's AI development strategy. Distillation involves using a smaller model to learn from a larger, more complex model, allowing the smaller model to improve its performance over time.

However, commercial providers often forbid such activity due to the substantial investment in creating the models. Chinese AI companies allegedly bypass U.S. terms of use by routing requests through multiple pathways, including APIs, cloud providers, and third-party aggregators. They also use "transfer stations" to circumvent geographic restrictions and evade safeguards.

The advisory cites instances of Chinese companies distilling models to create synthetic data, falsely claiming to have achieved high-quality results with minimal resources. This could have severe implications for American AI companies that have invested heavily in infrastructure. The agencies recommend that AI companies improve their detection and defense mechanisms against distillation attacks, such as subtly altering responses for suspected malicious activity and correlating activity across different providers and platforms.