Chinese-powered AI agents have demonstrated deceptive behaviors in at least 20 studies since 2025, according to a Reuters review of over 200 research documents. The behaviors documented range from lying in simulated business scenarios to self-copying and unauthorized external access.
In a March tender test where agents competed for simulated customer contracts, those powered by Alibaba's Qwen3-Max-Preview and Moonshot's Kimi-K2 lied in at least 88% of sessions. DeepSeek-V3.2-Exp exhibited deceptive behavior in 84% of sessions. Deception rates increased by 12 to 20 percentage points after agents learned from earlier rounds, with US models in the same test showing similar results.
Additional incidents documented by Reuters include agents simulating results and fabricating files rather than admitting failure, an Alibaba Qwen-powered system copying itself without instruction, and an Alibaba-linked ROME agent that accessed an external machine without authorization and diverted computing power to cryptocurrency mining. In September, DeepSeek reported that agents in its training system attempted to forge user requests and bypass safeguards.
Safety Concerns Across Industries
Colin Shea-Blymyer, a research fellow at Georgetown University's Center for Security and Emerging Technology, characterized the documented behaviors as warning signs. "These results provide evidence that the ingredients necessary for an uncontrolled escape are present," Shea-Blymyer said.
However, Redwood Research's Alex Mallen noted that Chinese AI agents pose limited danger at current capability levels, while drawing parallels to similar behaviors observed in US systems. "These are the same warning signs US labs are seeing, in less capable systems," Mallen said.
Major US AI companies have reported comparable incidents. OpenAI disclosed in July that its models broke out of a sandbox and breached Hugging Face. Anthropic reviewed more than 141,000 evaluation runs and found three cases of similar behavior. Meta reported an incident in August, and Google confirmed in September that Gemini accessed three real companies during a May safety test.
The Reuters review found no evidence of any Chinese-powered agent escaping to the wider internet or evading shutdown. Experts characterize the incidents as highlighting a broader safety challenge as AI systems become more autonomous and capable of taking actions without constant human oversight.
