Anthropic says new anti-distillation defenses cut data harvesting by up to 99%
Anthropic said on Monday that it has rolled out new safeguards aimed at stopping competitors from replicating Claude's capabilities, claiming the measures cut the number of reasoning tokens distillation attackers can extract to one-tenth to one-hundredth of previous levels — a roughly 90% to 99% reduction in harvested data. According to Techmeme's summary, the methods include "preserved thinking" techniques used in Fable 5.1 and Opus 5.5. Anthropic also said it has banned tens of thousands of accounts tied to distillation activity and pointed to Zhipu AI, saying it attempted to distill capabilities from Claude while developing its GLM-5.3 model, warning that such activity carries security and national-security risks.
If the defenses hold up under outside scrutiny, they could raise the cost and time needed to replicate frontier model capabilities cheaply — though the figures cited so far come from Anthropic itself.