Anthropic identifies over 23 million unauthorized Claude exchanges by Moonshot AI from May to July
The US developer claims Moonshot AI secretly routed nearly 300,000 customer requests through 5,380 fake accounts, feeding almost 23 million conversations to its Claude models.
Anthropic accused Beijing-based Moonshot AI of diverting user traffic intended for its Kimi chatbot to process via Anthropic's proprietary Claude models without disclosure. The US firm stated it discovered a ten-day window where approximately 300,000 requests passed through a network of 5,380 fraudulent accounts specifically targeting the high-performance Claude Opus model. Over three months spanning May to July, Moonshot facilitated more than 23 million interactions within the system. Anthropic alleges this operation utilized intermediaries and massive account generation to bypass regional restrictions and security detection mechanisms. The accusations are detailed in a threat intelligence report covering activities detected between December 2025 and August 2026. Additionally, the company reported that multiple other China-based AI laboratories attempted similar unauthorized extraction of capabilities through distillation or live traffic routing techniques.
