Anthropic claims new defenses cut AI distillation data theft by over 90%

Anthropic published a blog post on October 12 stating that its new anti-distillation defenses for Claude models have reduced the amount of data harvested by attackers by approximately 90 to 99 percent.

The company noted that while early measures like summarizing reasoning were effective, attackers shifted to active extraction techniques. In response, Anthropic introduced 'preserved thinking' in Claude Fable 5.1 and Opus 5.5, which prevents new API accounts from editing prior context, thereby blocking attempts to repeatedly extract hidden reasoning via proxy networks.

Additionally, reasoning transcripts are now accessible only to the organization whose request produced them. Anthropic acknowledged that these measures, while successful, resulted in turning away significant paid traffic from distillers and required some legitimate users to update their API integrations.