Anthropic Says Anti-Distillation Safeguards Cut Extraction by 90% or More

What happened

Anthropic announced on Monday, October 12, that new measures designed to stop rivals from copying Claude's capabilities through "distillation attacks" have reduced the data harvested by attackers by about 90% to 99%. The company says the safeguards restrict access to Claude's reasoning, bringing the number of reasoning tokens extracted by distillers down to between one-tenth and one-hundredth of previous levels, while tens of thousands of accounts linked to distillation campaigns have been banned.

Anthropic named Chinese AI company Z.ai (HKEX 2513.HK, also known as Zhipu), saying it had attempted to distill Claude's capabilities while developing its GLM-5.3 model, and warned that such practices pose safety and national security risks. Z.ai had not responded to a request for comment as of publication, Reuters reported.

Key facts

1. Timing: Anthropic announced on October 12; Reuters reported the news at 01:07 UTC on October 13.
2. Effect: Data harvested in distillation attacks down about 90%–99%; extracted reasoning tokens down to 1/10–1/100 of prior levels.
3. Methods: Restricted access to Claude's protected reasoning; cracked down on proxy networks repeatedly probing it, blocking large-scale distillation.
4. Bans: Tens of thousands of accounts linked to distillation campaigns banned.
5. Robustness: Anthropic says the safety system is "roughly as robust as those that prevent Claude from helping build bioweapons," with a false positive rate below 0.1%.
6. Named company: Z.ai (2513.HK, Zhipu) — cited for attempting to distill Claude's capabilities while developing GLM-5.3; Anthropic warns distillation poses safety and national security risks.
7. Context: Anthropic, OpenAI and Google all argue distillation lets rivals reproduce advanced capabilities faster and more cheaply; intensifying US-China AI competition has made it a shared concern among leading labs.