OpenAI and Anthropic Alert U.S. Government: Chinese AI Models Compromising Data Security and Rapid Advancement

AI leader Anthropic has accused China’s top labs—DeepSeek, MiniMax, and Moonshot AI—of illegally utilizing its Claude model outputs for training their systems. Business Insider reports that Anthropic claims these labs launched large-scale “distillation campaigns,” creating around 24,000 fraudulent Claude accounts that generated over 16 million unauthorized exchanges, violating service terms and regulations. Distillation, wherein smaller AI models are trained on larger models’ outputs, is common in AI development; however, Anthropic alleges these rivals exploit this method to steal technology rapidly and inexpensively. This warning aligns with similar concerns previously raised by OpenAI regarding DeepSeek’s practices and highlights the rising sophistication of distillation attacks. Anthropic CEO Dario Amodei emphasizes the need for U.S. export controls on advanced chips to curb misuse risks, like the development of bioweapons. Anthropic has introduced behavioral fingerprinting and collaborates with other firms to enhance defenses against these escalating threats.

Source link