Anthropic details Claude defenses against knowledge distillation
What happened
Anthropic published defenses that harden Claude against knowledge distillation attacks, Hacker News reported on October 13. The measures limit access to reasoning records across models and organizations, use a two-stage classifier, and restrict or ban accounts that repeatedly extract reasoning. Early estimates put the drop in data attackers can collect at roughly 90% to 99%, with the classifier's false positive rate under 0.1%. Some developers must update integrations, and Anthropic charges for blocked requests in certain domains. The company does not yet know how long the defenses will hold, and has not yet seen highly sophisticated, persistent bypass attempts.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageStrengthening our defenses against distillation attacks
Anthropic 公布 Claude 的抗知识蒸馏攻击方法，初步估计新防御使攻击者采集的数据减少约 90% 至 99%。措施包括限制推理记录跨模型和组织访问、采用两阶段分类器，以及限制和封禁反复提取推理的账户，分类器误报率低于 0.1%。这些措施要求部分开发者更新集成，并对特定领域被拦截的请求收费；Anthropic 尚不清楚效果能持续多久，也尚未遇到高度复杂且持续的绕过尝试。
Heat over time
Not enough continuous observations to draw a trend yet.