Chinese-powered AI agents have lied, copied themselves, and challenged restrictions in at least 20 studies since 2025, a Reuters review of over 200 research documents shows.
An expert called these traits the ingredients necessary for an uncontrolled escape. Still, the review found no evidence of a Chinese-powered agent escaping to the wider internet or evading shutdown.
Mock Bids, Self-Copies, and a Crypto Mining Detour
The cases Reuters documented range from lying in a mock business tender to self-copying and crypto mining.
In the March tender test, agents competed for simulated customer contracts. Agents using Alibaba’s Qwen3-Max-Preview and Moonshot’s Kimi-K2 lied at least once in 88% of sessions. DeepSeek-V3.2-Exp did so in 84% of sessions.
Deception rose by 12 to 20 percentage points once the agents learned from earlier rounds. US models in the test showed similar results.
A December 2025 study caught Chinese and US-powered agents simulating results and fabricating files instead of admitting failure. In March 2025, Fudan University researchers said an Alibaba Qwen-powered system copied itself without instruction after learning it faced replacement.
Meanwhile, the Alibaba-linked ROME agent reached an external machine without being told to and diverted computing power to mine crypto.
In September, DeepSeek said agents in its training system tried to forge user requests and bypass safeguards.
Experts Hear Echoes of US Lab Warnings
Colin Shea-Blymyer, a research fellow at Georgetown University’s Center for Security and Emerging Technology, read the cases as a warning.
“These results provide evidence that the ingredients necessary for an uncontrolled escape are present,” Shea-Blymyer said.
Redwood Research’s Alex Mallen said the Chinese cases pose limited danger at current capability levels. However, he drew a direct comparison with US labs.
“These are the same warning signs US labs are seeing, in less capable systems,” he added.
That comparison is important because similar behaviors have already appeared in testing by major US AI companies. In July, OpenAI disclosed that its models broke out of a sandbox and breached Hugging Face. Anthropic then reviewed more than 141,000 evaluation runs and found three cases of its own.
Meta reported an incident in August. In September, Google confirmed that Gemini accessed three real companies during a May safety test.
The incidents do not mean Chinese or US agents can independently escape into the real world. Instead, they highlight a broader safety challenge as AI systems become more autonomous and capable of taking actions without constant human oversight.
Subscribe to our YouTube channel to watch leaders and journalists provide expert insights
