How Foreign Actors Steal AI Models Through Distillation

We investigate how foreign threat actors, including Chinese AI labs, are illegally accessing and distilling proprietary AI models from Anthropic, Google, and OpenAI. The report details the scale of the problem, the methods used, and the security challenges facing the industry.

■ The Distillation Threat
- How stolen accounts and credit cards enable mass access to models like Claude
- Chinese labs (DeepSeek, MoonShot, MiniMax) accused of over 16 million interactions
- Distillation as a technique: training smaller models using larger "teacher" models

■ Security and Business Challenges
- Anthropic's response: account verification, fraud detection, and shutdowns
- The inherent trade-off between open access and security
- Economic impact: billions in investment, competition from cheaper copies

■ National Security Concerns
- Potential misuse for surveillance, cyberattacks, and bioweapons research
- U.S. government attention and proposed measures

This report is for anyone interested in AI security, intellectual property protection, and the geopolitical race for AI dominance. You will learn how model distillation works, why it is hard to stop, and what companies are doing to protect their technology.