techend
By Tanveer Ahmed :
Anthropic, an American artificial intelligence company, has publicly accused three Chinese technology firms of systematically harvesting capabilities from its Claude chatbot through the use of tens of thousands of fraudulent accounts and millions of interactions.
The San Francisco-based start-up alleges that DeepSeek, Moonshot AI, and MiniMax employed a technique called distillation—using outputs from a more advanced AI system to enhance the performance of their own less sophisticated models—on an industrial scale.
According to the company’s findings, the three Chinese laboratories collectively generated approximately 16 million exchanges with Claude, routing their activities through proxy services to bypass Anthropic’s ban on commercial access from China. The operation reportedly involved around 24,000 fake accounts.
Scale and Focus of Alleged Operations
Anthropic’s investigation reveals that each targeted company concentrated on different aspects of Claude’s capabilities. MiniMax allegedly conducted the most extensive campaign, generating more than 13 million interactions focused on coding, agentic reasoning, and tool use—areas where Claude is considered a market leader.
Moonshot AI purportedly generated over 3.4 million exchanges targeting reasoning and computer-use capabilities, while DeepSeek conducted approximately 150,000 interactions aimed at extracting step-by-step reasoning processes.
The company warned that such practices are growing in both intensity and sophistication, emphasizing that the window for effective countermeasures is narrowing. Anthropic called for coordinated action from industry and government, arguing that no single company can address the challenge alone.
National Security Dimensions
Beyond commercial concerns, Anthropic framed the issue as carrying significant national security implications. The company argued that models built through illicit distillation may not retain critical safety guardrails designed to prevent misuse, including restrictions on assisting with bioweapons development or enabling cyberattacks.
The statement also raised concerns about foreign laboratories incorporating unprotected model behaviour into military, intelligence, and surveillance applications, warning that open-sourcing such models could spread those risks more broadly.
Industry Reactions and Criticism
The accusations have generated substantial debate within the technology community, with some prominent figures questioning the moral authority of Anthropic’s position.
Elon Musk, founder of competing AI firm xAI, responded sharply on social media, noting that Anthropic recently paid $1.5 billion to settle a lawsuit with authors and publishers over using copyrighted books without permission to train its models. He characterized this as theft of training data at massive scale.
Software engineer Gergely Orosz made similar observations, pointing out that Claude’s training relied on copyrighted materials until litigation forced payment to rights holders, suggesting the company cannot have it both ways on intellectual property questions.
Broader Context of US-China Technology Competition
The allegations arrive amid heightened scrutiny of Chinese AI development in Washington. Earlier this month, OpenAI made comparable accusations to US lawmakers, describing ongoing efforts by Chinese companies to free-ride on capabilities developed by American frontier laboratories.
A Trump administration official separately told Reuters that DeepSeek likely used Nvidia’s advanced Blackwell chips—subject to US export controls—to train its upcoming V4 model, intensifying concerns about technology transfer.
China’s embassy in Washington responded to these broader tensions by stating that Beijing opposes drawing ideological lines, overstretching national security concepts, and abusing export controls to politicize economic, trade, and technological issues.
Legal and Ethical Questions
The controversy highlights fundamental questions about intellectual property boundaries in the AI era. Distillation exists in a legal grey area, with many models trained on publicly available internet data that may include outputs from other AI systems.
Observers have noted instances where Anthropic’s own models have inadvertently identified themselves as DeepSeek, illustrating the complex ways AI-generated content circulates through the open web and becomes incorporated into training data.
None of the three Chinese companies have issued formal responses to Anthropic’s allegations at the time of writing.
About The Author
