Anthropic announced it has significantly reduced the theft of Claude's reasoning tokens by up to 99% in its efforts to combat industrial-scale distillation. The company said its new defenses have sharply curtailed the volume of extracted reasoning tokens, the internal chain-of-thought steps of its models, from attackers.
According to Anthropic, attackers previously used networks of hundreds of thousands of fake accounts to siphon Claude's capabilities. The company claims that successful distillation attempts dropped sharply after its defenses went live, though some legitimate users have been mistakenly banned in the process.
Anthropic named Chinese companies Zhipu and Moonshot as participants in these attacks. Zhipu allegedly siphoned cyber capabilities from leading US models for its GLM-5.3 model, while Moonshot secretly routed queries through Claude. The stolen capabilities could end up in foreign military and surveillance systems.
"We have implemented a classifier that detects extraction attempts and automatically bans flagged accounts," said Anthropic's security team. The company added that tens of thousands of accounts have already been banned as part of its ongoing efforts.
The announcement follows growing concerns over AI model theft and the potential misuse of stolen capabilities in foreign systems. Anthropic is calling on other AI developers to adopt similar protections, though it acknowledges the challenge of balancing security with user access.
Anthropic did not say how many accounts remain active or how many legitimate users have been affected. The company also raised the question of how to effectively distinguish between malicious and benign behavior without impacting genuine users. The next step involves refining the system to minimize false positives.
Source: thedecoder
