Anthropic reports progress against distillation attacks

發生什麼事

Anthropic said on X that over the past few months it has significantly reduced how well distillation attacks work against Claude (@AnthropicAI). In a distillation attack, large numbers of fake accounts query Claude for answers and the outputs are used to train another model. According to Anthropic's article, the new approach has three layers: limiting where reasoning content can be carried, using a two-stage classifier to block requests that try to extract full reasoning, and rate-limiting and banning accounts that keep trying. Official figures: the classifier's false-positive rate is under 0.1%; the reasoning tokens distillers obtain on Anthropic's platform have fallen to between one-tenth and one-hundredth of previous levels; and tens of thousands of related accounts have been banned (Anthropic's article). Anthropic also acknowledged the costs: some developers have to update their integrations, and blocked requests are now still billed.