Anthropic has successfully reduced Claude reasoning token theft by up to 99 percent amid an ongoing battle against industrial-scale model distillation. The company reported that malicious actors have been systematically copying AI model capabilities using networks comprising hundreds of thousands of fake accounts. These attacks aim to extract internal chain-of-thought steps to train competing models.
As part of its findings, Anthropic specifically named Chinese companies Zhipu and Moonshot as participants in these extraction efforts. According to the report, Zhipu allegedly siphoned cyber capabilities for its GLM-5.3 model, while Moonshot secretly routed queries through Claude. The company warned that these stolen capabilities could potentially end up integrated into foreign military and surveillance systems.
To combat these threats, Anthropic implemented new defenses that restrict access to Claude's internal reasoning steps and deployed a classifier to detect extraction attempts, which has already resulted in tens of thousands of account bans. The company is now urging other artificial intelligence developers to adopt similar security protections, despite acknowledging that some legitimate users are inadvertently affected by the stricter controls.
Why it matters: This defense highlights the escalating security challenges artificial intelligence labs face in protecting proprietary reasoning capabilities from industrial-scale theft.