Anthropic says it cut distillation theft by up to 99 percent, and names Zhipu

The news

Anthropic published details of three layers of protection against distillation, training a rival model on a stronger model's outputs: limiting where Claude's reasoning text can go, a two-stage classifier that blocks extraction requests, and charges for blocked requests plus usage limits and bans for repeat offenders. It says extracted reasoning tokens are now 10 to 100 times lower than before.

Why it matters

Distillation is the cheapest way to close the gap with the frontier. This is the first detailed public account of a lab fighting it at scale, and it names names.

The details

The post accuses Zhipu of distilling cyber capabilities from leading US models while building GLM-5.3, and says requests to Moonshot's Kimi models were secretly forwarded to Claude. Zhipu did not respond to Reuters' request for comment.

Anthropic says it has banned tens of thousands of accounts with a classifier false-positive rate below 0.1 percent, calls the results early evidence, and admits its goal of zero impact on legitimate users was not met.

What we could not check

• We cannot independently verify the attribution to Zhipu or Moonshot.
• Detection and false-positive rates are Anthropic's own.