AIAnthropic says new defences cut the Claude data copied by rival labs by up to 99%
- Data harvested
- -90 to 99%
- early evidence
- Accounts banned
- Tens of thousands
- Classifier false positives
- <0.1%
Distillation attacks pull a strong model's outputs at scale, often through fake accounts, to train a cheaper copy. Anthropic said on 12 October that new defences had cut the amount of Claude's reasoning harvested this way by around 90 to 99%.
The measures include stopping new API accounts from editing Claude's earlier reasoning and a classifier that flags extraction attempts. Anthropic said it has banned tens of thousands of accounts, often less than two hours after a campaign starts.
Anthropic named Chinese lab Zhipu, saying it distilled cyber skills from leading US models while building GLM-5.3. The company said the defences cost it meaningful revenue and that some developers will need to change their integrations.
