Anthropic says it is making it much harder for other AI companies to copy Claude. According to a post by tech journalist Abhishek Yadav on October 13, 2026, the company says its new protections now block around 90 to 99 percent of attempts to copy its model through a technique called distillation. Anthropic is the company behind this site’s AI assistant, so we are keeping to what was reported and noting where claims are the company’s own.
What Was Reported?
The post says some companies use thousands of fake accounts to ask Claude huge numbers of questions, record how it responds, and train their own cheaper AI on those answers. Anthropic says its new protections block around 90 to 99 percent of such attempts, and accounts caught copying are banned, often within about 2 hours. The post adds that Anthropic says China’s Zhipu used Claude, along with other US models, to help build its GLM-5.3 model. Learn about the technique on the knowledge distillation Wikipedia page and see Anthropic’s news at anthropic.com/news.
Anthropic’s Claims at a Glance
| Item | Reported detail |
|---|---|
| Threat | Distillation using thousands of fake accounts |
| Block rate | Around 90-99% of attempts, per Anthropic |
| Enforcement | Accounts caught copying are banned, often within 2 hours |
| Named case | Zhipu’s GLM-5.3 allegedly built with help from Claude and other US models |
| Source of claims | Anthropic, as relayed in the post |
What Is Distillation?
Distillation means training a smaller or cheaper model on the outputs of a stronger one. It is widely used legitimately, for example by companies distilling their own large models into smaller ones. The controversy arises when someone does it to a rival’s model without permission, often at large scale, which most AI companies’ terms of service forbid. Follow our Anthropic coverage for related news.
Why It Matters
Frontier models cost a great deal to build, so copying their behaviour cheaply is attractive to competitors. Companies respond with detection systems that spot unusual traffic, such as thousands of coordinated accounts. If Anthropic’s numbers hold, that would raise the cost of copying. The mention of Zhipu also touches on wider tensions around US and Chinese AI development.
What Is Not Confirmed
These points are not confirmed by independent sources:
- The 90-99% block rate, which is Anthropic’s own claim.
- The 2-hour ban speed, also from the company.
- The claim about Zhipu and GLM-5.3, for which we saw no response from Zhipu.
- How the protections work, since details were not shared in the post.
- Whether other labs face the same problem at the same scale.
Quick Answers
What did Anthropic say? That new protections block about 90-99% of copying attempts.
What is distillation? Training a cheaper model on another model’s answers.
What happens to copying accounts? They are banned, often within about 2 hours.
Who was named? Zhipu, over its GLM-5.3 model.
Source: Post by Abhishek Yadav (@yabhishekhd) on X, October 13, 2026. Claims are Anthropic’s as relayed there; we did not see a response from Zhipu.
The wider context is a long-running debate over how AI labs should protect their models. Because models respond to any user who can pay for access, a determined copier can gather huge amounts of output through ordinary-looking requests. Detection often relies on patterns such as many accounts sending similar prompts, unusual volumes or sign-ups that look automated. The cat-and-mouse nature of this contest means no company can claim perfect protection, which is why the quoted 90 to 99 percent range is a claim worth watching for independent confirmation. Policymakers have also begun to look at the issue as part of broader discussions on AI competition and export controls, though no specific new rule was mentioned in the post.
