Anthropic's New Defenses Cut Distillation Attacks by Up to 99%
Strengthening our defenses against distillation attacks
Anthropic has rolled out stronger defenses against distillation attacks—covert campaigns that extract a model's reasoning to train rivals. Early results show a 90–99% drop in data harvested. The measures include limiting where reasoning can go, blocking extraction attempts with classifiers, and making repeated tries harder. The company has turned away trillions of tokens of paying traffic and banned tens of thousands of fraudulent accounts, though some legitimate users face integration updates and blocked-request charges.
Distillation enables companies to build models that approach the capability frontier, but which have, in practice, been released without proportionally robust safeguards.