Anthropic: "Blocked Up to 99% of Tech Theft Attempts"... Directly Targets China's Zhipu

Anthropic Announces "New Security System"
AI Inference Data Leaks Reduced by 90% to 99%
Urges Industrywide Action to "Stop Illicit Distillation"

U.S. artificial intelligence (AI) company Anthropic said it has introduced a new security system to prevent technology theft by foreign AI companies, including those in China, reducing the amount of AI inference data leaked by up to 99%. It also called for a coordinated industry response to stop "illicit distillation," saying Chinese companies are using advanced U.S. AI models without authorization to reduce development costs.

Anthropic CEO Dario Amodei speaks at an AI event on Feb. 19 (local time). Reuters-Yonhap News

In a blog post on Oct. 12 (local time), Anthropic said a new defense technology designed to block "distillation attacks" targeting its Claude AI model had reduced the amount of data collected by attackers by approximately 90% to 99%.

Distillation involves obtaining large amounts of responses and inference processes from a high-performing AI model and using them to train another model.

AI distillation itself is not illegal, but when carried out without the original model developer's permission, it can be considered a means of stealing technology.

Anthropic said, "Since we introduced the new defense system, the number of inference tokens extracted by attackers has fallen by 10 to 100 times," adding that it had blocked tens of thousands of accounts involved in distillation attacks. It said some accounts were detected within two hours of launching an attack.

The company also alleged that Chinese AI firm Zhipu had extracted the cybersecurity capabilities of major U.S. AI models without authorization while developing its AI model GLM-5.3. In particular, Anthropic said that after Claude's safeguards interfered with the technology extraction, the attackers shifted their targets to earlier versions of Claude and AI models from other companies.

Anthropic warned that AI models developed through unauthorized distillation could be misused for cyberattacks if released without safeguards. It alleged, for example, that Zhipu's GLM-5.3 could easily bypass security measures and be used to develop hacking tools.

Anthropic also raised concerns about national security. Among the requests sent to Claude through Kimi, an AI model developed by China's Moonshot, was one from a user presumed to be affiliated with the People's Liberation Army who asked for analysis of CCTV surveillance footage of a specific individual.

After attackers bypassed its security network using hundreds of thousands of fake accounts, Anthropic restricted access to AI inference processes, introduced attack-detection systems, and blocked accounts involved in repeated attacks.

Anthropic said it had given up substantial revenue and paid requests amounting to trillions of tokens in the course of these countermeasures. Nevertheless, it called on other AI companies and cloud platforms to coordinate their efforts, saying, "Preventing illicit distillation requires the effort and participation of every developer of cutting-edge AI models."

This content was produced with the assistance of AI translation services.