Anthropic blocks 99 percent of AI model theft and names Chinese rival
Rivals are stealing AI capabilities by asking clever questions. Anthropic just deployed a defense that cuts this theft by up to 99 percent.
Models · Source: Reuters
What happened
Anthropic just revealed aggressive new defenses to stop rivals from copying Claude. Competitors use a tactic called model distillation to steal advanced capabilities. They ask Claude thousands of clever questions and use the resulting answers to train their own cheaper models. Anthropic says its new safeguards have sharply reduced this theft. The data harvested in these distillation attacks has dropped by 90 to 99 percent.
The company made it much harder to use proxy networks to repeatedly probe Claude for its protected reasoning data. By blocking these extraction attempts, the amount of reasoning tokens extracted by distillers is now 10 to 100 times lower than before. To enforce this strict perimeter, Anthropic banned tens of thousands of accounts linked to organized distillation campaigns. They are actively hindering large scale efforts to clone their technology.
Anthropic directly named the Chinese AI startup Z.ai, also known as Zhipu. They accused Z.ai of attempting to distill capabilities from Claude while developing its own GLM-5.3 model. Anthropic, along with OpenAI and Google, argues this practice is a massive problem due to intensifying competition between the US and China. They warn that letting rivals reproduce advanced capabilities quickly and cheaply poses severe safety and national security risks.
Key facts
- 90 to 99 percent — Reduction in data harvested from distillation attacks
- 10 to 100 times — Decrease in reasoning tokens extracted by distillers
- Less than 0.1 percent — False positive rate of the new safety system
- Tens of thousands — Accounts banned for links to distillation campaigns
- Z.ai — Chinese startup accused of distilling Claude for its GLM-5.3 model
Why it matters
If you build AI products, you must understand that the era of cheap model cloning is rapidly ending. Startups and international rivals used to rely on querying top tier models to train their own specialized systems. Leading labs are now treating this extraction as a direct threat to their survival. They are actively hunting and banning accounts that look like they are scraping reasoning data for training purposes. The free ride of using frontier models as teachers is over.
The second order effect is a massive shift in how AI companies view and defend their moats. Anthropic says its anti distillation system is roughly as robust as the safeguards that prevent Claude from helping users build bioweapons. That comparison shows exactly how seriously they take this threat. Protecting the model weights is no longer enough to secure a competitive advantage. The new moat is strict, algorithmic control over who gets to ask questions, what proxy networks they use, and what kind of reasoning they are allowed to extract.
For builders
Risk of sudden account bans
If your application generates massive amounts of synthetic data, you might trigger Anthropic defenses. They boast a false positive rate of less than 0.1 percent for this safety system. However, getting caught in a massive ban wave will instantly break your product and cut off your access.
Higher costs for custom models
Distillation was the secret weapon for building cheap specialized models without massive research budgets. With frontier labs aggressively blocking extraction attempts, you will have to change your strategy. You must now pay for legitimate fine tuning or spend heavily to generate original training data from scratch.
Increased scrutiny on data pipelines
Labs are now actively monitoring proxy networks and repeated probing attempts. You can no longer hide behind fake accounts to scrape reasoning tokens. Builders must ensure their data collection methods comply with strict terms of service or risk losing access to foundational models entirely.
My take
We are learning that the real AI moat is not the model itself, but the permission to use it. Anthropic treating data extraction with the same severity as bioweapon prevention shows how desperate labs are to protect their billions in research. If your entire startup relies on scraping the brains of frontier models to build your own, your business model just got destroyed. You do not own the AI, you just rent the permission to ask it questions.
Original reporting: Reuters. This is my rewrite and opinion.