Anthropic has detailed new defenses against distillation attacks, covert campaigns that copy Claude's capabilities to build rival models without permission.

The measures restrict access to Claude's reasoning transcripts, the data attackers value most, use a two-stage classifier to block attempts to extract them, and apply stricter usage limits to repeat offenders. Early evidence suggests the amount of data distillers can harvest has dropped by roughly 90 to 99%, and tens of thousands of linked accounts have been banned.