Anthropic tightens protections for Claude reasoning against distillation attacks
A two-stage classifier and preserved-thinking controls reduced extracted reasoning data by 90–99 percent in Anthropic’s measurements.
AT A GLANCE
Key points
- Extracted reasoning fell by a reported 10–100×
- A two-stage classifier protects generally available frontier models
- The measured false-positive rate is below 0.1%
- Anthropic says the gains may not be permanent
What is changing?
Anthropic described new defenses against unauthorized distillation of Claude models on October 12. Distillation attacks use large networks of fraudulent accounts to collect model responses and train another system at a fraction of the original development cost.
The new approach limits the movement of reasoning across conversations, accounts, and weaker models, blocks extraction attempts with a two-stage classifier, and restricts persistent accounts. Anthropic reports a false-positive rate below 0.1 percent for the classifier.
According to the company’s own platform data, reasoning tokens extracted by attackers fell by a factor of 10 to 100. Anthropic explicitly says this is early evidence rather than a guarantee of lasting success, and notes that preserved thinking required some developers to update their integrations.
Explore the original sources
Source date: 12 October 2026.
Published by AI Radar: 14 Oct 2026, 03:03 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…