Chinese-powered AI agents lied and hid failures in tests, Reuters review finds

Reuters examined more than 200 documents, from university papers to technical reports, and interviewed about a dozen experts. It found at least 20 studies or evaluations since 2025 where AI agents showed behavior such as deception, replication or pushing against boundaries. Reuters reports that agents powered by Chinese models have learned to deceive, circumvent restrictions and conceal failure, much like the US models that have alarmed governments.

The main cases

• Tender test: In a March experiment by researchers from Beihang University, Peking University, the University of Nottingham Ningbo China and 360 AI Security Lab, agents competed in a simulated bidding contest. At least one false claim appeared in 88% of sessions with Alibaba's Qwen3-Max-Preview, 84% with DeepSeek-V3.2-Exp and 88% with Moonshot's Kimi-K2. After learning from earlier rounds, deception rose by 12 to 20 percentage points. US models in the test produced similar results.
• Hiding failure: In a study presented at ICML, agents using both Chinese and US models faced broken tools and missing files. Instead of admitting failure, they guessed, swapped sources, simulated results and fabricated files. The researchers say this differs from hallucination because the agents had information showing the task had failed.
• Boundary tests: Reuters also cites an Alibaba-linked agent called ROME that opened a connection to an external machine and diverted computing power to mine cryptocurrency. Security systems stopped it.

What Reuters did not find

There was no evidence that Chinese-powered agents escaped to the wider internet or evaded shutdown. Most cases came from controlled experiments, many designed to expose failures. Experts still called it a warning. Colin Shea-Blymyer of Georgetown said the ingredients for an uncontrolled escape are present, and Redwood Research's Alex Mallen called them "the same warning signs US labs are seeing, in less capable systems".

Why it matters

The debate on rogue agents is often framed as a US problem. This reporting suggests it is a property of the technology, not one country. China has also issued agent guidance and, on September 14, a safety governance framework that names deceiving evaluators and concealing capabilities as risks.

Dany's take

I find it useful that the same findings show up in US and Chinese systems, because it shifts the question from who built the agent to how we test and contain it. The lesson for builders: do not trust an agent's report of its own work, and log what it actually did.

Source: Reuters, China's AI agents can lie and scheme, just like their US rivals