Anthropic has accused explosive allegations against three of China’s leading artificial intelligence (AI) developers — DeepSeek, Moonshot AI, and MiniMax — of orchestrating a massive “distillation” campaign to illicitly siphon the capabilities of its flagship model, Claude.
In a detailed report released Monday, Anthropic claims the three companies engaged in approximately 16 million exchanges using a network of 24,000 fraudulent accounts. The goal was to “distill” Claude’s high-level reasoning, coding, and tool-use abilities into their own proprietary models — a move Anthropic claims violates both its terms of service and U.S. export controls.
Distillation is a common machine learning technique where a smaller model is trained on the outputs of a larger, more sophisticated teacher model. While Anthropic admits the practice is a legitimate tool for companies to create cheaper versions of their own software, it labeled the Chinese efforts as “adversarial.”
The San Francisco firm detailed a sophisticated hydra cluster strategy, in which the perpetrators allegedly used commercial proxy services to spread traffic across thousands of accounts, masking their activity to avoid detection. By flooding Claude with repetitive, highly specific prompts, the Chinese firms were reportedly able to extract chain-of-thought data, the step-by-step internal reasoning that makes frontier models powerful.
In response to the breaches, Anthropic says it is deploying “behavioral fingerprinting” to identify extraction patterns and is calling for a coordinated industry-wide defense.
The scale of the operation varied among the accused, according to Anthropic. MiniMax, it asserts, was the largest offender with more than 13 million exchanges. Anthropic noted that when a new version of Claude was released, MiniMax redirected half its traffic within 24 hours to capture the latest capabilities.
Moonshot AI, meanwhile, accounted for 3.4 million exchanges, focusing on agentic reasoning and computer vision, Anthropic said.
DeepSeek, it alleges, generated 150,000 exchanges focused on complex reasoning and censorship-safe rewrites of sensitive political queries.
DeepSeek previously sent shockwaves through Silicon Valley in early 2025 by releasing a high-performing model at a fraction of the traditional training cost, leading to a temporary $1 trillion market cap wipeout of U.S. tech giants.
The allegations come amid a deepening AI arms race between Washington, D.C., and Beijing. Anthropic framed the data extraction as a national security threat, warning that illicit distillation allows “authoritarian governments” to enhance military and surveillance systems using American innovation.
However, the tech community’s reaction has been polarized. Critics on social media and industry observers were quick to point out the irony of Anthropic’s stance. The company has faced its own legal battles — including a $1.5 billion settlement over the use of pirated books — leading some to view these allegations as a move to protect market share rather than national security. Elon Musk weighed in on X, mockingly rebranding the firm “MisAnthropic.”
