Anthropic said on October 12, 2026, its new AI safeguards reduced data harvesting from Claude by approximately 90% to 99%. The California-based AI company introduced stronger controls to stop rival developers from copying Claude's advanced capabilities through repeated model queries.
The measures target AI distillation attacks, which use outputs from advanced models to train competing systems more cheaply. Anthropic reported that extracted reasoning tokens are now 10 to 100 times lower than previous levels. The company also linked large-scale extraction attempts to intellectual property concerns and potential national security risks.
Anthropic has tightened access to Claude's internal reasoning and made automated extraction more difficult. New controls restrict attempts to manipulate conversation histories and expose protected reasoning through repeated prompts. The company also introduced reasoning summaries that reduce the amount of directly reusable information in model responses.
Additional safeguards target proxy networks that attackers use to send large volumes of requests. Anthropic has also banned tens of thousands of accounts associated with suspected distillation campaigns. The company said its detection system maintains a false positive rate below 0.1%.
Anthropic described the system as “roughly as robust as those that prevent Claude from helping build bioweapons.” The comparison highlights the company's emphasis on preventing advanced AI capabilities from reaching unauthorized users.
Anthropic identified Chinese AI startup Z.ai, also known as Zhipu, as an example of suspected distillation activity. The company alleged that Z.ai attempted to extract Claude's capabilities while developing its GLM-5.3 model. Z.ai did not immediately respond to a request for comment.
The allegations arrive as US and Chinese AI developers compete to build advanced models. Anthropic, OpenAI and Google have raised concerns that distillation allows competitors to replicate sophisticated capabilities with fewer resources.
Anthropic's figures represent the company's own measurements and have not received independent verification in the supplied report. The reported reduction signals stronger protection, although it does not establish that future extraction attempts will fail.
The company may still face attackers who adapt their methods as safeguards improve. Stronger access controls could nevertheless raise the cost of copying advanced AI capabilities and reduce unauthorized extraction.
Also Read: HCLTech and Anthropic Join Forces to Bring Agentic AI to Global Businesses