Anthropic Says New Defenses Cut AI Copycat Data Theft by 90%+

Anthropic announced on October 12 that new safeguards cut the reasoning data copycats can pull out of Claude by an estimated 90% to 99%. Distillation means training a cheaper model on a stronger model’s answers, and Anthropic is focused on groups doing it at industrial scale without permission.

Anthropic says it relies on three things: limits on how Claude’s reasoning can be reused across conversations and accounts, a two-stage classifier that flags extraction attempts, and tougher enforcement. It says tens of thousands of accounts have been banned, often within two hours of a campaign starting.

Why the reasoning matters: in Anthropic’s test, a model trained on full reasoning transcripts scored 80.4% on a coding benchmark, versus 71.8% when trained on summaries.

The catch: The 90% to 99% figure is Anthropic’s own estimate and covers only its own platform. Some developers need to update integrations. Accounts created on or after August 31, 2026 get the stricter context checks by default. Since September 24, requests blocked by the safeguards are billed, and Anthropic puts the false-positive rate below 0.1%. It also admits it does not know how long these gains will last.

Try it: If you use the Claude API, check your logs for new errors on requests that edit earlier reasoning.