Anthropic 指控中國公司大規模蒸餾 Claude

🔑 增強重點摘要

• •Anthropic published a detailed blog post on February 24, 2026, disclosing the distillation campaigns and attributing them to specific labs using request metadata, IP correlations, and infrastructure indicators[1][2][3].
• •Anthropic has implemented classifiers, behavioral fingerprinting, strengthened verification for educational and startup accounts, and enhanced safeguards to reduce model outputs' utility for distillation[2][3].
• •The incident follows Google's recent disruption of similar distillation attacks targeting Gemini's reasoning via over 100,000 prompts, as reported by Google Threat Intelligence Group[2].
• •Anthropic supports stricter US export controls on advanced chips and AI services, arguing that distillation relies on such hardware and undermines national security efforts[1][3].

🛠️ 技術深入

• •Distillation involves training a smaller, less capable model on outputs from a stronger frontier model like Claude to replicate capabilities at lower cost[1][2].
• •Attackers used 'hydra cluster' proxy architectures with tens of thousands of fraudulent accounts distributing traffic across API keys and cloud providers to evade bans and detection[1][3].
• •Anthropic's defenses include classifiers for suspicious API patterns, behavioral fingerprinting, and modifications to model outputs that hinder effective distillation training[2][3].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線