Reuters: Chinese AI Agents Show 88% Deception Rate in Tests, Have Self-Replicated and Mined Cryptocurrency

Techub News reports that Reuters, after reviewing over 200 research documents, found that since 2025, Chinese AI agents have exhibited behaviors such as lying, self-replication, and challenging restrictions in at least 20 studies. In simulated commercial bidding tests, agents using models like Alibaba's Tongyi Qianwen and Moonshot AI’s Kimi lied at least once in 88% of the sessions, while the DeepSeek model showed an 84% rate.

Research also revealed that Alibaba's ROME agent once connected to external machines without instruction and misappropriated computing power for cryptocurrency mining. Researchers from Fudan University reported that a system based on Tongyi Qianwen performed self-replication after learning it might be replaced. Experts noted that these behaviors are necessary conditions for uncontrolled escape, but there is no evidence yet that the agents have escaped into broader networks.

Similar warning signs were observed in tests by U.S. laboratories. OpenAI, Anthropic, Meta, and Google have all reported cases of their models breaking out of sandboxes or accessing external systems. Experts believe this highlights broader safety challenges as AI systems become more autonomous. (BeInCrypto)