Tag
推文提及 Anthropic 加强针对蒸馏攻击的防御措施，同时顺带提到 Bun 运行时的相关动态。
This paper argues that existing defenses against distillation attacks in large language models are ineffective when attackers further train their models with reinforcement learning, highlighting potential flaws in current threat models.
Anthropic accused Alibaba of creating tens of thousands of fake Claude accounts for distillation attacks, leading to stricter guardrails that are affecting legitimate users.
This article warns that current and upcoming AI models significantly lower the barrier to creating bioweapons, citing distillation attacks on open-weight models and the inability to prevent safety ablation. It calls for public funding of broad-spectrum countermeasures as a necessary response.
This paper studies distillation attacks where model outputs can enable imitation, proposing a minimax game framework and a forward-pass-only defense called Product-of-Experts, showing that adaptive students recover more capability than passive evaluation suggests.
Anthropic announced new defenses against illicit distillation attacks on Claude, claiming a 90-99% reduction in harvested data, while warning that distilled models like Zhipu's GLM-5.3 pose cybersecurity and national security risks.
Anthropic宣布在抵御针对Claude的蒸馏攻击方面取得重大进展，早期数据显示已将蒸馏者窃取的数据量减少约90%至99%。文章将非法蒸馏描述为国家安全风险，指出Zhipu在开发GLM-5.3时曾试图蒸馏Claude的网络攻击能力，且部分蒸馏模型的能力已被整合进外国军事和情报系统。