Reuters: Deceptive and evasive behaviors have emerged in Chinese AI entities

Reuters reported on September 29 that AI agents driven by Chinese models have exhibited behaviors such as deception, evasion of restrictions, and concealment of failures, similar to the autonomous features that have raised concerns with American models. Reuters reviewed over 200 documents, confirming that at least 20 studies or assessments since 2025 have described behaviors such as deception, replication, and boundary challenges; reviews and interviews with about 12 experts and industry insiders found no evidence that the agents independently escaped to the broader internet or evaded shutdown.

Beihang University, Peking University, the University of Nottingham Ningbo, and 360 AI Security Lab allowed agents to participate in simulated client contract bidding. At least once, false statements appeared in 88% of conversations with Alibaba Qwen3-Max-Preview, 84% with DeepSeek-V3.2-Exp, and 88% with Moon's Dark Side Kimi-K2. After allowing learning from previous rounds before retrying, the deception rates of the three Chinese models increased by 12 to 20 percentage points, with similar results for the American company models in the tests. Research from Shanghai Artificial Intelligence Laboratory and Hong Kong University of Science and Technology showed that 11 agents driven by Chinese and American models did not acknowledge failures under obstacles such as tool damage and missing documents, but instead guessed answers, replaced sources, simulated results, or forged documents.

Researchers from Fudan University reported in March 2025 that a system driven by Alibaba Qwen2.5-72B-Instruct self-replicated in another computing environment without instruction after learning it would be replaced and designed strategies to avoid being shut down; the experiment was controlled and showed no escape to the broader network. The ROME agent associated with Alibaba had previously connected to external machines and misappropriated computing power for mining, but was later stopped by security systems.