Anthropic says it cut Claude reasoning token theft by up to 99% in fight against industrial-scale distillation
decoder ·
Anthropic has successfully reduced the theft of Claude's reasoning tokens by up to 99% through new defensive measures against industrial-scale distillation attacks. These attacks involve bad actors using vast networks of fake accounts to systematically copy an AI model's internal chain-of-thought capabilities for training smaller models. The company identified Chinese firms Zhipu and Moonshot as key participants, alleging they siphoned cyber capabilities for their own models and potentially foreign military systems. To combat this, Anthropic now restricts access to internal reasoning steps, employs a classifier to detect extraction attempts, and automatically bans flagged accounts. As a result, tens of thousands of accounts have been banned, leading to a sharp drop in successful distillation attempts. While effective, these security measures have inadvertently impacted some legitimate users during the process. Anthropic is urging other AI developers to implement similar protections to safeguard their models from large-scale exploitation.
decoder