The Most Valuable Thing an AI Owns Is How It Thinks
Rivals can copy an AI's hard-won skills by quietly querying it millions of times. Anthropic says it now blocks most of that, though only its own numbers back the claim.
The smartest thing an AI company owns may not be its chips or its data. It may be the way its model thinks. On October 12, Anthropic said new defenses have cut how much of Claude's reasoning copycats can grab by about 90 to 99 percent.
The trick it is fighting is called distillation. A rival sends a big model millions of questions and saves the answers along with the step by step thinking behind them. Then it uses that pile to train a cheaper clone that mimics how the original reasons. Anthropic says these campaigns run through networks of fake accounts, and it names Chinese labs such as Zhipu and Moonshot. Those accusations are Anthropic's own, and Zhipu did not respond to Reuters.
Locking the Thinking
Anthropic says it now keeps Claude's reasoning tied to the account that asked for it. A screening system spots requests that try to pull out hidden thinking, and repeat offenders face limits or bans. The company reports banning tens of thousands of accounts, often within two hours.
The 90 to 99 percent figure comes only from Anthropic's own estimate of what was taken from its own platform. No one outside the company has checked it, and Anthropic admits it has not yet faced a truly determined attempt to beat the new defenses. Some honest developers have also had to change their code, and a small share of normal requests now get blocked.
Skills Are the New Border
For builders, the lesson is that a model's reasoning is an asset worth guarding, just like source code. Watch for odd patterns of use, not just bad words.
For countries, the stakes are bigger. Skills that take years and billions to build can leak out through an API in weeks. Sovereign AI means protecting hard-won capability and building your own, since copied skills often arrive without the original safety guards.
Sources
- Anthropic, Strengthening our defenses against distillation attacks, October 12, 2026
- GovInfoSecurity, Anthropic Rolls Out Defenses to Block AI Model Distillation, October 13, 2026
- Reuters via The Economic Times, Anthropic says new safeguards sharply reduce AI distillation attacks, October 13, 2026
- TechCrunch, Anthropic details distillation campaigns from Alibaba, Moonshot AI, and DeepSeek, September 10, 2026