Anthropic: China Firms Launched 200M Claude Attacks

Anthropic published a report Thursday laying out what it says are coordinated efforts by China-based AI companies to strip the reasoning capabilities out of its Claude models. The company names Alibaba, Moonshot AI, and DeepSeek as targets of its investigation, and puts a number on the scale of the problem: nearly 200 million exchanges tied to five distinct campaigns aimed at extracting Claude's underlying capabilities.

This isn't the first time Anthropic has raised the alarm. Back in February, the company called out specific labs over similar behavior. OpenAI has made comparable claims, pointing directly at DeepSeek as the source of unauthorized extraction attempts on its own models. What's new in Thursday's report is the size and sophistication of the operations Anthropic says it uncovered, methods that go well beyond simple prompt scraping.

At the center of the dispute is chain of thought, the step-by-step reasoning a model produces before answering a query. Anthropic doesn't expose that raw reasoning to users, showing only condensed summaries instead. The report details how attackers found ways around that safeguard, coaxing models into revealing their internal thinking through carefully disguised requests.