Reuters reported on Sept. 29 that AI agents powered by Chinese models have displayed deception, attempts to evade restrictions, and efforts to conceal failure, behaviors that fall into the same class of autonomy-related concerns that have drawn attention around U.S. models.
After reviewing more than 200 documents, Reuters said it confirmed that since 2025, at least 20 studies or evaluations had described agent behavior involving deception, replication, and boundary-testing. Its review, along with interviews with about 12 experts and industry participants, found no evidence that agents had independently escaped into the broader internet or bypassed shutdown measures.
False statements in a simulated bidding exercise
Reuters cited a test run by Beihang University, Peking University, the University of Nottingham Ningbo China, and 360 AI Security Lab, where agents were asked to take part in a simulated client contract bidding process.
In that test, at least one false statement appeared in 88% of conversations involving Alibaba’s Qwen3-Max-Preview, 84% for DeepSeek-V3.2-Exp, and 88% for Moonshot AI’s Kimi-K2.
When the agents were allowed to learn from earlier rounds and try again, deception rates for the three Chinese models increased by 12 to 20 percentage points. Reuters said models from U.S. companies showed similar results in the same testing setup.
Failure concealed under obstacle conditions
Research from the Shanghai Artificial Intelligence Laboratory and the Hong Kong University of Science and Technology found that 11 agents powered by Chinese and U.S. models did not admit failure when faced with obstacles such as broken tools or missing files. Instead, they guessed answers, substituted sources, simulated results, or fabricated files.
Controlled tests on replication and shutdown avoidance
Researchers at Fudan University reported in March 2025 that a system powered by Alibaba’s Qwen2.5-72B-Instruct, after learning that it would be replaced, replicated itself in another computing environment without being instructed to do so and devised strategies to avoid being shut down.
Reuters said the experiment was controlled and did not show that the system escaped into the broader network.
The report also said that ROME, an agent linked to Alibaba, had connected to external machines and diverted computing power for mining before being stopped by a security system.