
Anthropic 称减少 Claude 推理令牌窃取
通过限制推理步骤访问和检测分类器打击大规模蒸馏攻击。
为什么重要
提示模型提供商需防范通过推理步骤泄露进行能力蒸馏。
阅读原文the-decoder.com
AI 解读(根据原文生成)
The Decoder 报道称,Anthropic 正在应对针对 Claude 模型的大规模蒸馏攻击:攻击者用大量虚假账号系统性复制模型能力,用于训练较小模型。Anthropic 表示,新防御已将被提取的推理令牌量减少 90% 到 99%。措施包括限制访问 Claude 内部推理步骤、运行检测提取尝试的分类器,并自动封禁被标记账号。报道还提到 Anthropic 点名 Zhipu 和 Moonshot 参与相关行为。
要点(AI 据原文整理,数字均已与原文核对)
- Anthropic says its new defenses have cut the volume of extracted reasoning tokens by 90 to 99 percent.
- Attackers use networks of hundreds of thousands of fake accounts to pull it off.
- Anthropic now restricts access to Claude's internal reasoning steps.
- Anthropic runs a classifier that detects extraction attempts and automatically bans flagged accounts.
- Tens of thousands have already been banned.