Anthropic details its anti-distillation defenses
The lab describes controls on reasoning, classifiers and measures against repeated extraction attempts.
publication
Facts, analysis and limits
Verified facts
- A 90 to 99 per cent reduction in data collected by distillers.
- Some context/reasoning measures were introduced with September models.
- Legitimate integrations may require adaptation.
What changes
Blocks of reasoning become more strongly related to their model and context.
Practical implication
Respect replay constraints and do not depend on hidden reasoning text.
Availability
Belgium/EEA: not specified in the source; access depends on the plan, rollout and administrator settings.
Limits and uncertainty
Report published on October 12, not simultaneous activation of all measures; figures and attributions are Anthropic's.