IT Home, October 13: On Monday local time, Anthropic announced the introduction of several new protective measures designed to prevent competitors from replicating the Claude AI model. These measures have reduced the amount of data stolen through model distillation attacks by approximately 90% to 99%.
Model distillation has become a major concern for leading AI companies, with Anthropic, OpenAI, Google, and others believing that this technique allows competitors to replicate cutting-edge model capabilities at lower cost and faster speed.
Anthropic stated that this set of protective mechanisms designed to restrict external access to Claude's reasoning process has significantly reduced information extraction attempts. Now, "the number of reasoning tokens that distilled attackers can extract is only one-tenth to one-hundredth of what it was previously."
IT Home noted that the California-based company stated the new protection makes it more difficult to repeatedly probe Claude’s protected inference pathways via proxy networks, blocks attempts at information extraction, and hinders large-scale distillation efforts.
Anthropic stated that the robustness of this security system is “roughly equivalent to the safeguards preventing Claude from assisting in the creation of biological weapons,” with a false positive rate below 0.1%.
Anthropic has banned tens of thousands of accounts associated with distillation activities, citing the Chinese AI startup Zhipu AI as an example, stating that the company attempted to extract model capabilities from Claude during the development of its GLM-5.3 model. Anthropic warned that such activities pose security and national security risks.