Anthropic says new safeguards cut model-distillation extraction 90–99%

What changed

Anthropic’s October 12 research post details defenses against attempts to extract a frontier model’s hidden reasoning and use it to train a competing model. The company says its updated safeguards cut the reasoning tokens available to distillers by a factor of 10 to 100, with some days’ extracted amount too small to measure. It estimates the total harvested data fell around 90–99%.

The system combines controls on how reasoning can be moved between models, accounts and conversations with a two-stage classifier trained against extraction attempts. Anthropic says the classifier has a false-positive rate below 0.1% and that it has banned tens of thousands of accounts associated with distillation campaigns. These numbers are company measurements, and the firm describes the work as an ongoing cat-and-mouse process.

What the result means

The trade-off is practical: preserved-thinking changes have required some developers to update integrations, and a small number of benign requests may be blocked or charged. This is a frontier-model security story as much as a product story: model providers are trying to protect training advantages without making legitimate API use brittle.