MockingbirdNews Logo

Mockingbird News

REAL NEWS NEVER FELT FUNNIER

Categories

Chinese AI Labs Use Claude for Training, Pretend It's Their Homework

Chinese AI Labs Use Claude for Training, Pretend It's Their Homework
Photo by Jon Tyson on Unsplash

KEY POINTS

  • Anthropic accused Alibaba, Moonshot AI, DeepSeek, Zhipu, and Xiaomi of launching nearly 190 million AI distillation attacks on Claude from May to July 2026.
  • Alibaba’s network of 3,500 fraudulent accounts alone ran over 151 million exchanges, siphoning Claude’s reasoning to train their Qwen models.
  • In response, Anthropic enhanced Claude’s security by limiting detailed reasoning outputs and requiring identity checks from users linked to China, Russia, or Iran.

Between May and July 2026, China’s AI giants – Alibaba, Moonshot, DeepSeek, Zhipu, and Xiaomi – ran a covert summer camp for AI freeloading by launching nearly 190 million distillation attacks on Anthropic's Claude. Alibaba led the pack with over 151 million sneaky chats from 3,500 fake accounts, milking Claude’s reasoning like a cash cow for their Qwen models. Moonshot confessed to rerouting 23 million user queries to Claude, while DeepSeek slid in 12.1 million shortcut requests in just half a month. Meanwhile, Xiaomi fed user chats from its MiMo bot right into Claude, apparently without telling anyone. Anthropic even caught Chinese and Russian military secrets broadcasted accidentally, thanks to mix-ups in AI needle-in-a-haystack routing. In response, Anthropic revamped Claude's security, making the bot less of an open diary and asking users from banned countries to prove they aren't bots dressed as humans. No one cared to comment, proving silence is the Chinese AI labs' best encryption method.

Share the Story

(1 of 3)

Source: Businessinsider | Published: 9/11/2026 | Author: Aditi Bharade

Read the original article →