Anthropic has implemented new security measures to stop companies from using fake accounts to distill and copy the internal reasoning capabilities of its Claude AI models.
Key facts
- •Anthropic reduced the volume of extracted reasoning tokens by 90 to 99 percent.
- •Attackers utilize networks consisting of hundreds of thousands of fake accounts.
- •Anthropic alleges Zhipu used stolen data for its GLM-5.3 model.
- •Moonshot is accused of secretly routing queries through Claude.
- •Anthropic has banned tens of thousands of accounts as part of its new security strategy.
Anthropic has successfully reduced the unauthorized extraction of Claude's internal reasoning tokens by 90 to 99 percent. The company implemented new defensive measures to combat industrial-scale distillation attacks, where entities use networks of hundreds of thousands of fake accounts to copy the model's capabilities.
Allegations against Chinese firms
Anthropic has identified Chinese companies Zhipu and Moonshot as participants in these extraction efforts. The company alleges that Zhipu siphoned cyber capabilities from U.S. models to develop its GLM-5.3 model, while Moonshot reportedly routed queries through Claude to extract data. Anthropic stated that these stolen capabilities could potentially be integrated into foreign military and surveillance systems.
New defensive measures
To prevent further theft, Anthropic now restricts access to Claude's internal chain-of-thought steps and utilizes a classifier to detect and block extraction attempts. The company has already banned tens of thousands of accounts identified by these systems. Anthropic acknowledged that these automated defenses have occasionally impacted legitimate users and is encouraging other AI developers to adopt similar security protocols.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by The Decoder.
