Anthropic Alleges Chinese AI Labs Ran Massive Campaign to Clone Claude

In a detailed report released Monday, AI developer Anthropic stated it uncovered what it calls "industrial-scale campaigns" by three Chinese artificial intelligence firms to copy the core capabilities of its Claude model. The company alleges that DeepSeek, Moonshot AI, and MiniMax orchestrated the effort, generating over 16 million queries through roughly 24,000 fraudulent accounts.

According to Anthropic, the campaigns violated its terms of service and regional access restrictions, which prohibit use in China due to legal and security concerns. The technique involved is known as distillation, where a less advanced model is trained on outputs from a superior one. While a legitimate practice for companies refining their own models, Anthropic contends it is illegal when used by competitors to rapidly acquire technology developed elsewhere.

"Illicitly distilled models lack necessary safeguards, creating significant national security risks," the company warned. It argued that such copied models could strip away safety protections, allowing dangerous capabilities to proliferate and potentially be weaponized for cyber operations, disinformation, or surveillance by foreign governments.

The attacks, Anthropic says, were sophisticated. They relied on commercial proxy services using vast networks of accounts to mask the activity. One network managed over 20,000 accounts simultaneously. The prompts were not random; they specifically targeted Claude's most advanced functions like reasoning, coding, and tool use across millions of exchanges.

In response, Anthropic has deployed new detection systems, strengthened account verification, and implemented safeguards to reduce the usefulness of its outputs for such extraction. This disclosure follows a similar report from Google last month concerning attacks on its Gemini model, highlighting a growing concern within the industry about protecting proprietary AI technology.

Source: The Hackers News