Anthropic Accuses Chinese AI Labs of Large-Scale Claude Data Mining Amid U.S. Chip Export Debate

Washington D.C. — In a startling revelation that has intensified the global AI security debate, U.S.-based artificial intelligence firm Anthropic has formally accused three Chinese AI laboratories—DeepSeek, Moonshot, and MiniMax—of orchestrating a coordinated, large-scale data extraction campaign targeting its proprietary Claude models. According to internal forensic analyses, the labs allegedly deployed over 24,000 synthetic user accounts to interact with Claude’s public interfaces, systematically harvesting responses to probe model behavior, reasoning patterns, and safety guardrails. The operation, which spanned late 2025 through early 2026, appears designed to distill Claude’s advanced capabilities for use in training competing models, a practice known in AI circles as "model distillation" or "stealth fine-tuning."

Anthropic’s investigation, detailed in an internal security bulletin obtained by this outlet, revealed anomalous query patterns that deviated sharply from typical user behavior. The fake accounts exhibited unusually high volumes of complex, multi-turn prompts—often focused on coding, logical reasoning, and agent-based tasks—mirroring the exact capabilities highlighted in Anthropic’s January 2026 announcement of Claude Opus 4.6, its most powerful model to date. "This wasn’t accidental usage," said a senior security analyst at Anthropic, speaking on condition of anonymity. "It was a systematic, resource-intensive effort to reverse-engineer our architecture without direct access to weights or training data."

The timing of the allegations coincides with escalating U.S. government scrutiny over semiconductor exports to China. As the Biden administration weighs new restrictions on advanced AI chips—particularly high-bandwidth memory (HBM) and next-generation NVIDIA H200 and B200 units—Anthropic’s findings have been cited in closed-door briefings to the Committee on Foreign Investment in the United States (CFIUS) and the Department of Commerce. "This isn’t just intellectual property theft," said a senior U.S. tech policy official. "It’s a strategic effort to circumvent export controls by building AI models that don’t require the same level of hardware, using our models as training proxies."

Anthropic’s response has been multifaceted. In addition to blocking the identified IP ranges and deploying adversarial detection filters, the company launched Anthropic Labs in January 2026—a dedicated R&D unit focused on "ethical AI defense." The initiative, first announced in a company blog post, aims to develop countermeasures against model extraction attacks and to pioneer "privacy-preserving AI" techniques. "We’re not just building smarter models," said Anthropic’s Chief Scientist in a recent interview. "We’re building models that can defend themselves."

Meanwhile, the release of Claude Opus 4.6 on February 5, 2026, underscores the competitive urgency. With a 1-million-token context window and breakthrough performance in coding and autonomous agent workflows, Opus 4.6 represents the current pinnacle of Anthropic’s hybrid reasoning architecture. Industry analysts note that the model’s precision in long-horizon tasks makes it an exceptionally valuable target for distillation. "If you can mimic Opus 4.6’s reasoning without the hardware, you’ve effectively leapfrogged the export controls," said Dr. Elena Ruiz, AI policy fellow at the Center for Security and Emerging Technology.

Chinese AI firms have not publicly responded to the allegations. DeepSeek and Moonshot continue to release open-weight models with claims of "native innovation," while MiniMax has not commented. However, independent researchers have noted striking similarities in reasoning patterns between Claude Opus 4.5/4.6 and recently released Chinese models, particularly in handling ambiguous prompts and multi-step code generation.

As the U.S. Congress prepares to vote on expanded AI chip export controls this month, Anthropic’s allegations have injected new urgency into a debate once focused on hardware. Now, the question is no longer just how to limit access to chips—but how to protect the intellectual output they enable. The global AI race has entered a new phase: one where models are not just tools, but targets.