They Stole AI With 24,000 Fake Accounts

Quick Overview

Anthropic revealed that three AI labs—DeepSeek, Moonshot, and MiniMax—conducted industrial-scale distillation attacks, using 24,000 fake accounts to acquire the chain-of-thought reasoning and output pairs from Anthropic's Opus model without permission, which is a violation of terms of service and may be illegal under existing copyright law, prompting Anthropic to publicly expose the activity.

Key Points: Anthropic identified industrial-scale distillation attacks targeting their Claude model by three AI labs: DeepSeek, Moonshot, and MiniMax. The attacks used approximately 24,000 fraudulent accounts and proxy services to extract chain-of-thought training data from the Opus teacher model. DeepSeek executed over 150,000 exchanges targeting reasoning capabilities and creating censorship-safe alternatives to policy-sensitive queries. Moonshot AI conducted over 3.4 million exchanges targeting agentic reasoning, coding, data analysis, computer-use agent development, and computer vision. MiniMax executed the largest attack with over 13 million exchanges, focusing on agentic coding and tool use/orchestration. Anthropic claims this data extraction method is essentially stealing the sophisticated reasoning traces developed by the high-cost, US-based frontier models, circumventing export controls aimed at preventing Chinese labs from accessing advanced compute capabilities. The company settled a separate copyright infringement lawsuit with authors for $1.5 billion, paying $3,000 per estimated 500,000 books used in training, highlighting the contentious legal landscape around training data.

Context: The video discusses Anthropic's recent blog post detailing coordinated, industrial-scale "distillation attacks" targeting their large language model, Opus. Model distillation is a technique where a smaller 'student' model learns from the outputs (including internal reasoning or chain-of-thought) of a larger, more capable 'teacher' model. Anthropic exposed that three specific labs—DeepSeek, Moonshot, and MiniMax—were engaging in this process using thousands of fraudulent accounts to illicitly gain access to Opus's proprietary reasoning capabilities, which is costly to develop and potentially violates export controls.