Anthropic 偵測 DeepSeek、Moonshot、MiniMax 蒸餾 Claude

🔑 增強重點摘要

• •The campaigns generated over 16 million exchanges using approximately 24,000 fraudulent accounts, violating Anthropic's terms and China access ban[1][2][3].
• •DeepSeek's campaign involved over 150,000 exchanges focused on reasoning across diverse tasks, using synchronized traffic and shared payment methods for load balancing[1][3].
• •Moonshot AI conducted over 3.4 million exchanges targeting agentic reasoning, tool use, coding, data analysis, computer-use agents, and computer vision to reconstruct reasoning traces[1][3].
• •MiniMax executed the largest campaign with over 13 million exchanges on agentic coding and tool use, pivoting nearly half its traffic to a new Claude model within 24 hours of release[1][2][4].

🛠️ 技術深入

• •Distillation campaigns used 'hydra cluster' architectures: commercial proxy services distributing requests across thousands of fraudulent accounts, mixing distillation queries with mundane ones to evade detection[1][4].
• •Attribution relied on IP address correlation, request metadata, infrastructure indicators, and industry partner corroboration matching actor behaviors on other platforms[2][3].
• •DeepSeek targeted censorship-safe query rewrites, prompting Claude to rephrase sensitive political topics for training models to bypass safety filters[4].
• •Anthropic deployed classifiers, behavioral fingerprinting for API traffic, strengthened educational/startup verifications, and output safeguards to reduce distillation efficacy[2][3].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線