Anthropic says new defenses cut data extraction from Claude by distillers by 90% to 99%

Anthropic said on Monday, October 12, that its new measures against "distillation" cut the data attackers could pull from Claude by 90% to 99%, according to Reuters.

What is distillation?

It is a practice where a group uses the answers of an advanced model to train a smaller one that imitates it, usually at lower cost. Reuters explains it has become a key concern among leading AI companies amid US-China competition.

What Anthropic says

• The defenses focus on limiting access to Claude's internal reasoning. According to Reuters, reasoning tokens extracted are now 10 to 100 times lower than before.
• It says it made it harder to use proxy networks to repeatedly probe the system.
• Its system has a false positive rate of less than 0.1%.
• It banned tens of thousands of accounts linked to distillation campaigns.
• It named Chinese company Z.ai (Zhipu) as an example: it allegedly tried to distill cyber capabilities from Claude while developing its GLM-5.3 model.
• The company also says it has turned away meaningful revenue from traffic tied to this practice.

Why it matters to you

For creators and businesses using Claude through the app or API, the measures carry trade-offs that Anthropic itself acknowledges, according to Pulse 2.0:

• Some developers had to update their integrations because of a feature called "preserved thinking."
• Legitimate requests can occasionally be blocked.
• Users may be charged for requests the system rejects. Anthropic says it resumed charging for blocked requests in certain categories, including distillation, where its classifiers have false-positive rates below 0.1%.

All figures are Anthropic's claims; the sources do not include independent verification. Z.ai did not immediately respond to Reuters' request for comment.

Sources

• Reuters via Yahoo: new safeguards against distillation
• Pulse 2.0: defenses cut 90% to 99%