Western AI labs are collaborating to combat distillation attacks, where smaller models are trained using outputs from more advanced systems. This practice allows developers to emulate the capabilities of frontier models at a lower cost, which major firms argue undermines their significant investments in AI development.
Anthropic reported on various distillation tactics in September 2026, noting that attackers often use engineered prompts to extract internal reasoning traces. In response, the company has implemented measures such as banning accounts, blocking IP addresses, and requiring models to summarize reasoning before providing full responses.
The issue is complicated by the purchase of third-party chat logs and the use of gray market transfer stations to bypass geo-restrictions. While CISA has issued recommendations for detection and prevention, no official legislation sanctioning these activities has been enacted to date.
The ongoing tension highlights a divide between the proprietary approach of U.S. companies and the open-weight models favored by many Chinese developers. The topic is expected to be part of the discussions during the meeting between President Trump and Premier Xi Jinping on September 24.
