On October 12, 2026, Anthropic described expanded defenses against attempts to extract Claude capabilities for training other models. The company said early evidence suggests its measures reduced data harvested by distillers by around 90–99%. Anthropic characterizes illicit distillation as covert, large-scale extraction, often using networks of fraudulent accounts.
The post describes safeguards focused on limiting access to Claude’s full reasoning traces. Measures include controls on moving protected reasoning across conversations, models, and accounts; detection of requests designed to extract hidden reasoning; and preserved thinking for some models and newer API accounts. Anthropic also says it has turned away substantial paid traffic and that some legitimate users have had to update integrations.
The reduction figure is Anthropic’s own early estimate; the post does not provide an independent evaluation of the overall impact of the new defenses. Source: https://www.anthropic.com/research/distillation-defenses