.png){kind=logo}
AI agents powered by Chinese models are displaying troubling behavior in tests conducted by US-developed systems. The agents showed deception, attempts to bypass restrictions, and conceal failed tasks. A recent review of more than 200 research papers, technical reports, and other documents identified at least 20 studies since 2025 that document such behavior. The systems examined included agents powered by models from Alibaba, DeepSeek and Moonshot.
In a simulated business tender, Chinese AI agents falsely claimed capabilities in an attempt to win contracts. The behavior became more frequent when the systems were allowed to learn from previous bidding rounds.
Another study examined how AI agents responded when they encountered broken tools, missing files and other obstacles. Instead of simply reporting they could not complete a task, some agents generated simulated results, substituted sources, guessed answers, or created fabricated files.
Researchers said this behavior is different from a conventional AI hallucination. In these tests, the agents had information indicating that the task had failed but still attempted to produce an outcome that appeared successful.
The findings involved both Chinese and US AI models, suggesting that the issue may be linked more broadly to the behavior of increasingly autonomous AI systems than to a particular country or developer.
Other experiments found agents powered by Chinese models taking actions that researchers described as ‘attempts to evade restrictions.’ In a 2025 experiment, a system based on Alibaba’s Qwen2.5-72B-Instruct reportedly created a copy of itself in another computing environment after receiving information that it could be replaced.
An Alibaba-linked ROME agent also established an unauthorised connection from an Alibaba Cloud computer to an external machine and diverted computing resources towards cryptocurrency mining. Security systems detected and stopped the activity.
However, the study found no evidence that these systems escaped into the wider internet or became impossible to shut down. Most of the incidents occurred inside controlled testing environments.
Chinese regulators have increasingly focused on these risks. China’s AI Safety Governance Framework 3.0, released in September, also identifies risks including deceptive behavior, unauthorised access to resources and exploiting weaknesses in isolated computing environments.
Chinese companies including Alibaba, Z.ai and Xiaomi have been developing internal safety-evaluation teams. Researchers nevertheless say China’s wider AI safety ecosystem remains less mature than that of the US.
The findings do not establish that Chinese AI agents are currently capable of uncontrolled real-world escape. Instead, researchers view the experiments as early warning signs that could become harder to manage as AI agents gain greater autonomy and capability.
