Anthropic Disrupts IBM, Accuses Chinese Firms of IP Theft
Anthropic's enterprise impact was highlighted as IBM's shares dropped 13% following the announcement that its Claude model could be used for COBOL modernization. In a separate development, Anthropic has also accused Chinese companies of intellectual property theft, alleging they are using its API to distill and train their own large language models.
- The 13% drop in IBM's stock was its most significant single-day decline in over 25 years, erasing nearly $30 billion in market value. This was a reaction to the perceived threat that AI-driven COBOL modernization poses to IBM's substantial revenue from mainframe-related consulting and services. - The core of COBOL modernization with AI involves reverse-engineering business logic from systems that can be over 50 years old. Tools like Claude are used to map code dependencies, document workflows that are no longer understood by current employees, and automatically translate the logic into modern languages. - Anthropic has specifically named three Chinese AI firms—DeepSeek, Moonshot AI, and MiniMax—in its accusations of intellectual property theft. The alleged scheme involved creating approximately 24,000 fraudulent accounts to conduct over 16 million interactions with the Claude API. - The technique in question, "model distillation," involves using the outputs of a more powerful "teacher" model (like Claude) to train a smaller "student" model. This allows a competitor to replicate a model's capabilities at a fraction of the cost and time of developing it independently. - From an MLOps perspective, this incident highlights the critical need for robust API security to prevent model extraction. Best practices include implementing strict access controls, monitoring for anomalous query patterns, and employing techniques like API rate limiting and model watermarking. - The alleged IP theft was focused on areas where Claude is considered a leader, such as advanced coding, agentic reasoning, and tool use. To bypass regional restrictions, the firms allegedly routed their API traffic through third-party proxy services. - This is not an isolated incident; OpenAI has made similar accusations against Chinese companies, and Google has also observed model extraction attacks on its AI models, signaling a growing trend of AI-related intellectual property theft. - While the modernization of legacy code is the immediate application, the underlying AI capability is a form of automated program synthesis. Research in this area, often featured at conferences like NeurIPS, explores using techniques like reinforcement learning to improve the functional correctness of generated code.
