Anthropic says copying of Claude's capabilities is down by up to 99%

Anthropic — the US company behind the Claude AI assistant — announced on Monday (12 October 2026) that new security measures have sharply reduced so-called "distillation" attacks: data harvested by attackers is down by 90 to 99 percent, Reuters reports.

"Distillation" is a technique in which a rival company asks an AI system a huge number of questions — especially about its step-by-step reasoning — then uses the answers to train its own model. That way it can copy advanced capabilities quickly and cheaply, instead of spending billions of dollars training a system from scratch. Anthropic, OpenAI and Google say the practice has become a major concern, amid intensifying technology rivalry between the United States and China.

The new safeguards restrict access to Claude's "reasoning" — the thinking steps the model shows before answering. Anthropic says the number of "reasoning tokens" extracted by copyists is now 10 to 100 times lower than before. It has also become harder to use proxy networks to probe the system repeatedly — the company says it blocks extraction attempts and disrupts large-scale copying campaigns.

Anthropic says its defence system is "roughly as robust as" the systems that stop Claude from helping build bioweapons — and that its false-positive rate (wrongly blocking innocent users) is under 0.1%.

The company also says it has banned tens of thousands of accounts linked to copying campaigns, and named the Chinese startup Z.ai (also known as Zhipu) as an example: it says the company tried to distill Claude's capabilities while developing its GLM-5.3 model — warning such practices pose safety and national security risks. Z.ai did not immediately respond to Reuters' request for comment.

What does this mean for you?

For developers and tech entrepreneurs: If you build apps on top of Claude or GPT APIs, this fight shapes the economics of the tools you use. Cheap models from China are what many can afford — and now you know why the big labs are pushing back: they claim some rival models are built by copying their work. Choose your tools knowing where they come from.

For small businesses: AI tools keep getting cheaper as competition grows — but a low price can hide legal and ethical costs. Use tools from sources you trust, and remember that completely free services usually come with hidden terms.

For students: Training a top-tier AI system costs billions of dollars — which is why the best models are not free. "Distillation" is one way others try to skip those costs. Understanding this helps explain why companies guard their models so fiercely.

For policy watchers: This is a new front in the US–China technology rivalry — after chip restrictions and visa fees, the fight is now over who owns the "minds" themselves.