Anthropic于周四发布报告，揭示了阿里巴巴、Moonshot AI和DeepSeek等中国AI公司针对其Claude模型开展的持续蒸馏攻击活动1。这些攻击旨在提取Claude的思维链过程，用于训练规模更小的模型1。
根据报告，Anthropic共观察到近2亿次与蒸馏攻击相关的交互，来自五个独立活动1。其中阿里巴巴的活动规模最大，在2026年5月至7月期间记录了1.51亿次交互，日峰值接近300万次，涉及跨越3500个账户的请求1。Moonshot AI的活动规模较小，在十天内生成近30万个请求，通过5000个账户路由，其中部分请求来自中国军方1。
攻击者重点针对Claude的多项关键能力，包括代理能力、工具使用、编码、数据分析和逻辑推理等1。Anthropic曾在2月份就蒸馏攻击问题发表过声明1，此前OpenAI也报告过类似活动并将其指向DeepSeek1。
Anthropic disclosed Thursday a series of ongoing distillation attacks orchestrated by Chinese artificial intelligence companies, including Alibaba, Moonshot AI, and DeepSeek, aimed at extracting capabilities from its Claude model.1 The company documented approximately 200 million interactions connected to distillation attacks across five separate campaigns.1 These attacks seek to harvest Claude's chain-of-thought reasoning to train smaller, competing models.1
Alibaba's campaign represented the largest operation, with 151 million interactions logged between May and July 2026, peaking at nearly 3 million requests daily across approximately 3,500 accounts.1 Moonshot AI conducted a more concentrated effort, generating approximately 300,000 requests over ten days routed through 5,000 accounts, with some requests originating from Chinese military sources.1 The distillation attacks specifically targeted Claude's core capabilities, including agent functionality, tool usage, coding, data analysis, and logical reasoning.1 Anthropic had previously issued a statement on distillation attacks in February, while OpenAI similarly reported comparable activities and attributed them to DeepSeek.1
评论
还没有评论，欢迎留下第一条。