- Anthropic identified three Chinese AI firms using over 24,000 fraudulent accounts for industrial-scale “distillation attacks” on its Claude model.
- The illicit campaigns generated over 16 million exchanges to extract advanced capabilities like reasoning and coding, violating terms of service and regional bans.
- The company warns such distillation strips crucial safety safeguards, creating significant national security risks that authoritarian governments could weaponize.
- Google recently disclosed similar attacks on its Gemini model, indicating a growing trend of AI model extraction.
On February 24, 2026, AI company Anthropic publicly accused three Chinese competitors of massive, illegal campaigns to steal its technology. According to its reports, DeepSeek, Moonshot AI, and MiniMax orchestrated “industrial-scale campaigns” using fraudulent accounts to extract Claude‘s advanced capabilities.
These distillation attacks generated over 16 million exchanges through about 24,000 accounts, violating strict regional restrictions on its services in China. The technique involves training a weaker model on a stronger AI system’s outputs to cheaply acquire its skills.
However, illicit distillation bypasses the massive development costs rivals would normally face. Anthropic stated “Illicitly distilled models lack necessary safeguards, creating significant national security risks.”
Consequently, foreign entities could weaponize these unprotected capabilities for malicious cyber activities or surveillance systems. The campaigns specifically targeted Claude’s most advanced features like agentic reasoning and coding across millions of prompts.
Anthropic attributed each attack using metadata and infrastructure clues, noting the prompts’ volume and structure were distinct from normal use. Meanwhile, the attacks relied on proxy services with “hydra cluster” architectures to avoid detection.
To counter this, the company has built classifiers and strengthened verification for certain account types. This disclosure follows similar findings from Google, which recently disrupted attacks on its Gemini model.
✅ Follow BITNEWSBOT on Telegram, Facebook, LinkedIn, X.com, and Google News for instant updates.
