Anthropic Flags Large-Scale AI Model Theft Campaigns

Anthropic has disclosed what it describes as large-scale, coordinated attempts by three China-based AI firms—DeepSeek, Moonshot AI, and MiniMax—to extract capabilities from its Claude model through systematic “model distillation.” According to the company, the campaigns relied on roughly 24,000 fraudulent accounts and commercial proxy infrastructure to generate more than 16 million queries designed to replicate Claude’s advanced reasoning, coding, and agentic tool-use behaviors.

The attackers allegedly deployed “hydra cluster” architectures—distributed traffic systems that fragment high-volume requests across thousands of accounts to evade detection. Instead of obvious scraping, prompts were blended with legitimate-looking usage patterns, making the activity harder to distinguish from normal developer traffic.

The objective was not simple data harvesting, but capability replication. By collecting high-quality Claude outputs, the firms could fine-tune or train competing models at dramatically lower cost and time compared to building frontier capabilities independently.

Anthropic’s internal findings suggest differentiated targeting strategies:

• DeepSeek focused on advanced reasoning and politically sensitive content generation.

• Moonshot AI targeted agentic reasoning and computer-use capabilities.

• MiniMax concentrated heavily on advanced coding and tool orchestration workflows.

This incident underscores a growing strategic threat: model extraction as industrial espionage. Frontier AI development requires massive compute investment and proprietary safety engineering. Distilled replicas may lack embedded safeguards, increasing risks of misuse in surveillance, cyber operations, or information manipulation.

In response, Anthropic says it has strengthened behavioral fingerprinting, traffic anomaly detection, and classifier-based defenses to block coordinated distillation attempts.

The episode signals a broader shift in AI competition. As models become core national and commercial assets, protecting them from systematic capability theft is rapidly becoming as critical as building them.