Anthropic Accuses Chinese Labs of Using 24000 Fake Accounts to Siphon US AI Tech

The global race for artificial intelligence dominance has taken a sharp, contentious turn. Anthropic, the San Francisco-based creator of the Claude chatbot, has gone public with explosive allegations against three prominent Chinese AI laboratories. According to a detailed report first shared with federal authorities, the firms DeepSeek, Moonshot AI, and MiniMax allegedly orchestrated a massive, industrial-scale campaign to "siphon" proprietary intelligence from American systems.

The scale of the operation is staggering. Anthropic's threat intelligence team identified more than 24,000 fraudulent accounts used to generate over 16 million exchanges with Claude. The goal? A sophisticated technique known as "model distillation." In essence, these labs are accused of using Claude’s high-level reasoning and coding outputs to "teach" their own models, effectively leapfrogging years of research and billions in development costs.

The Mechanics of a Distillation Attack

While distillation is a common practice in the AI industry—often used by developers to create smaller, more efficient versions of their own models—Anthropic argues this was anything but standard. The company claims the Chinese labs used proxy services and "hydra clusters" of fake accounts to bypass regional restrictions, as Claude is not officially available for commercial use in China.

The siphoning was specifically targeted at Claude’s most advanced features, including:

• Complex Agentic Reasoning: Reconstructing the step-by-step "thought processes" Claude uses to solve problems.

• Advanced Programming: Extracting high-tier coding logic to bolster Chinese software-writing AI.

• Tool Orchestration: Learning how Claude interacts with external software and APIs to perform autonomous tasks.

According to Anthropic's official statement, MiniMax alone was responsible for roughly 13 million of these exchanges. Perhaps most concerning was the speed of the operation; when Anthropic released a new model update, MiniMax allegedly pivoted nearly half of its "attack traffic" to the new system within just 24 hours.

National Security and the Export Control Dilemma

This news comes at a time of heightened tension between Washington and Beijing over high-end semiconductors. While the U.S. has tightened export controls on Nvidia chips to slow China's AI progress, these allegations suggest that Chinese firms are finding digital backdoors to the same intelligence. By querying American models through thousands of fake identities, these labs can essentially "clone" the intelligence of restricted tech without ever needing the underlying hardware.

Anthropic warns that this poses a significant national security risk. "Foreign labs that distill American models can then feed these unprotected capabilities into military, intelligence, and surveillance systems," the company stated. They argue that because these distilled models are built without Anthropic's proprietary safety guardrails, they could be repurposed for offensive cyber operations or disinformation campaigns.

Industry Backlash and the Hypocrisy Debate

Despite the gravity of the claims, the news has been met with a wave of skepticism from tech industry veterans. Critics, including Elon Musk, have been quick to point out the irony. Musk took to X (formerly Twitter) to suggest that Anthropic itself was built on "stolen" data, referring to the company’s history of scraping the open internet to train Claude.

Furthermore, as reported by The Decoder, Anthropic recently settled a $1.5 billion lawsuit with authors and publishers over copyright infringement. This has led many in the developer community to question whether a company that trained on public data can ethically complain when others do the same via an API.

Regardless of the moral debate, the fallout is real. This coordinated siphoning effort underscores a new frontier in tech espionage—one where the "theft" isn't of hardware or code, but of the very logic and reasoning patterns that define modern AI.