Anthropic accuses Chinese AI labs of ‘industrial-scale’ theft to bypass US export curbs

Anthropic has leveled blistering accusations against three prominent Chinese artificial intelligence laboratories, alleging “industrial-scale” campaigns of intellectual property theft that the startup claims undermine US national security and circumvent rigorous export controls.

The San Francisco-based creator of the Claude AI interface asserted Monday that DeepSeek, Moonshot, and MiniMax have engaged in “industrial-scale distillation attacks” against its proprietary models.

Distillation refers to the practice of training smaller, leaner AI systems on the outputs of more sophisticated models. This technique allows developers to replicate elite performance levels without requiring the massive computational infrastructure typically necessary to achieve such benchmarks.

The issue has gained significant geopolitical resonance as Chinese AI groups grapple with tightening US export restrictions. These controls limit access to Nvidia’s most advanced hardware, including the high-end Blackwell series, forcing firms in the world’s second-largest economy to pivot toward alternative strategies. To mitigate the impact of the silicon blockade, Chinese entities have reportedly turned to offshore model training, the utilization of legacy or black-market semiconductors, and aggressive engineering efficiencies to drive down costs.

Anthropic disclosed it has detected 24,000 fraudulent accounts linked to the effort, documenting more than 16 million interactions with Claude. The company alleges these communications were systematically engineered to “train and refine” the Chinese firms’ internal models.

“Distillation attacks allow foreign laboratories, including those under the control of the Chinese Communist Party, to erode the competitive advantage that export controls were designed to protect, effectively neutralizing those safeguards through other means,” Anthropic stated.

Over the past month, Chinese AI groups including Moonshot and MiniMax have debuted a suite of high-performance models. These releases have earned praise from developers for their efficiency in powering AI agents and video generation applications.

DeepSeek, which stunned Silicon Valley in January with its breakthrough R1 model—celebrated for matching Western performance at a fraction of the cost—had been widely expected to unveil a new iteration ahead of last week’s Lunar New Year holiday. However, the company has yet to issue a formal announcement.

Since the debut of R1, DeepSeek has focused on incremental updates rather than a new flagship release. This strategic pause has allowed domestic rivals such as Alibaba and ByteDance to gain ground, attracting developers seeking low-cost, open-source alternatives.

US artificial intelligence firms and Washington officials have intensified warnings that distilled models present acute national security risks.

Anthropic argued that models “created through illicit distillation” are unlikely to retain the rigorous safety guardrails designed to prevent AI from being weaponized for biological warfare or malicious cyber operations.

The allegations follow a similar move by OpenAI in January. The ChatGPT creator revealed it had uncovered evidence that its foundational models were being distilled, pointing to DeepSeek as the suspected beneficiary.

Earlier this month, OpenAI submitted a memorandum to the US House Select Committee on Strategic Competition between the United States and the Chinese Communist Party. The memo argued that DeepSeek’s upcoming model must be viewed within the context of “efforts to leverage capabilities developed by OpenAI and other US frontier laboratories for free.”

The startup further accused DeepSeek of reflecting “CCP censorship and information control,” maintaining that “adversarial distillation poses serious cost, security, commercial, and strategic risks for the US.”