Anthropic Introduces New Defenses to Cut Model Distillation Data Theft by 99%

Anthropic has rolled out new security measures that reduce data theft from model distillation attacks by 90% to 99%. The company also banned tens of thousands of accounts linked to distillation activities. This development significantly increases the difficulty for competitors to replicate cutting-edge AI models at a lower cost and faster speed. It addresses a major industry concern regarding intellectual property protection and AI safety. The new security system restricts outside access to Claude's reasoning processes, keeping extracted reasoning tokens to just one-tenth to one-hundredth of previous levels. Anthropic reported that the system has a false positive rate of less than 0.1%.