Anthropic reported that it has significantly reduced the extraction of internal reasoning tokens from its Claude models, achieving a drop of 90 to 99 percent. According to reporting by The Decoder, the artificial intelligence company connected these high-volume distillation attempts to Chinese firms Zhipu and Moonshot. In response to the activity, Anthropic stated that it has banned tens of thousands of accounts involved in the industrial-scale harvesting operation.
The Threat of Model Distillation
Distillation is a technique where a smaller artificial intelligence model is trained using the outputs of a more advanced, larger model to mimic its capabilities at a lower cost. When developers manage to capture internal reasoning tokens, which represent the step-by-step logic an AI uses before answering a prompt, they gain valuable data that can accelerate the training of competing models without incurring the full research and development expense.
What Remains Unclear
Despite the company’s announcements regarding the blocked accounts and the percentage reduction in token theft, several details have not been made public. The announcement did not specify the exact technical methods Anthropic used to identify and intercept the extraction attempts, nor did it disclose the total volume of tokens stolen before the countermeasures were fully deployed. Furthermore, the public statements did not outline whether legal actions beyond account bans are being pursued against Zhipu and Moonshot.
As competition in the foundational model market intensifies, securing proprietary model outputs and internal architectures will likely remain a critical operational challenge for major artificial intelligence laboratories.
Lead image: Christina Morillo / Pexels
