Anthropic Slashes AI Model Theft Risk

Security firm Anthropic reduced industrial-scale AI model theft by 90-99% through targeted defenses against distillation attacks.

Anthropic reduced industrial-scale AI model theft by 90 to 99 percent after deploying defenses against distillation attacks—systematic copying of models to train smaller versions. The company implemented restrictions on Claude’s internal reasoning steps and a classifier to detect extraction attempts, which automatically banned tens of thousands of accounts involved in the attacks. Zhipu and Moonshot were identified as participants: Zhipu allegedly gained capabilities from leading US models for its GLM-5.3, while Moonshot routed queries through Claude without authorization. This directly strengthens security for advanced AI access by curbing industrial-scale model theft, though legitimate users were impacted by the defense measures. The reduction applies specifically to extracted reasoning tokens, not overall model integrity. The source provides limited detail on the scale of attacks or long-term effectiveness beyond the initial defense activation. This brief is an honest summary of the source’s reporting; deeper technical impacts and adoption rates sit with the original source.

Source: The Decoder