Anthropic Battles Alleged AI Model Theft: Chinese Firms Accused of Mining Claude

Anthropic, the AI start-up behind the Claude chatbot, has publicly accused three major Chinese AI firms—DeepSeek, Moonshot AI, and MiniMax—of orchestrating a large-scale effort to extract and replicate Claude’s advanced capabilities, igniting fresh tensions in the global AI race and raising national security concerns.

Key Takeaways

• Anthropic claims over 16 million queries were sent to Claude by more than 24,000 fake accounts linked to the Chinese firms.

• The alleged activity used a method called “distillation” to copy advanced reasoning, coding, and tool-use skills.

• The accusations surface amid escalating debates over US export controls on advanced AI chips to China.

The Allegations: Massive Distillation Campaign

According to Anthropic, the three Chinese companies leveraged thousands of fake accounts to send millions of queries to Claude, enabling them to systematically extract complex responses. This process, known as model distillation, typically trains a "student" AI model to mimic the outputs of a more advanced "teacher"—a widely-used technique for optimizing one’s own AI models, but highly controversial when targeting external technology.

Anthropic asserts that these attempts went far beyond routine usage patterns, instead targeting Claude’s most distinctive strengths: agentic reasoning, coding abilities, chain-of-thought outputs, and nuanced handling of sensitive topics. Each Chinese firm is believed to have focused on different aspects, with MiniMax responsible for the largest share at 13 million exchanges.

Industry Impact And Geopolitical Stakes

The revelations come at a critical moment when both industry leaders and lawmakers are grappling with how to police intellectual property in the age of generative AI and whether to tighten restrictions on exports of advanced chips vital for training such models. Anthropic’s claims followed similar accusations by OpenAI against DeepSeek and coincided with Google disclosing its own struggles to protect its AI from copycat attacks.

There’s concern that these illicitly trained models—often stripped of the safety features built into Western AIs—could be repurposed for malicious applications including cyberattacks, disinformation, or mass surveillance. Given Anthropic’s contracts with sensitive US government entities, including defense and intelligence, the stakes reportedly extend well beyond commercial competition.

Calls For Collective Action

Anthropic insists the industry cannot tackle the threat alone. It has implemented new detection measures, such as behavioral fingerprinting and stricter account verification, but the company urges a coordinated response from AI developers, cloud service providers, and policymakers. The findings have renewed calls for tighter control over AI model access and hardware exports, with some experts suggesting a complete halt to advanced chip sales to implicated overseas firms.

A Debate Of Ethics And Irony

While Anthropic’s claims highlight serious risks, industry observers have noted the irony—many leading US AI labs, including Anthropic itself, trained their own models with data scraped from the public internet, often without explicit consent. This has fueled a nuanced debate about intellectual property, competition, and ethical boundaries in AI development.

As DeepSeek prepares to launch its latest model, and with the US and China locked in an intensifying technological rivalry, the question of how to safeguard advanced AI models from large-scale replication remains both urgent and unresolved.

Sources

• Anthropic Says Chinese AI Firms Used 16 Million Claude Queries to Copy Model, The Hacker News.

• Anthropic Furious at DeepSeek for Copying Its AI Without Permission, Which Is Pretty Ironic When You ConsiderHow It Built Claude in the First Place, Futurism.

• Anthropic claims Chinese firms copied Claude model used by Pentagon via fake accounts, ynetnews.

• Anthropic accuses Chinese AI labs of mining Claude as US debates AI chip exports, TechCrunch.