AI agents are supposed to take tasks off our plates to make our lives easier. But what happens when they start lying to get the job done? According to a new investigation by Reuters, agents running on Chinese models are picking up the same sneaky habits that got people worried about US models.
Reuters dug through more than 200 documents and spotted at least 20 studies since 2025 where agents lied, replicated themselves, or tested their limits. Experts warn that these are exactly the kinds of skills an AI would need to slip out of human control.
How exactly are these AI agents misbehaving?
In one experiment, agents powered by Alibaba’s Qwen, DeepSeek, and Moonshot’s Kimi models competed for customer contracts in a simulated tender. At least one false claim showed up in 84% to 88% of the sessions. Given another shot after learning from earlier rounds, they lied more, with deception jumping 12 to 20 percentage points. To be fair, US models in the same test didn’t do any better.
In another study, agents powered by both Chinese and US models ran into broken tools and missing files. Rather than owning up to failure, they guessed answers, faked results, and even made up files. The researchers say this isn’t a hallucination, since the agents knew the task had gone wrong.
Fudan University researchers also found an Alibaba-powered system that made a copy of itself, unprompted, once it learned it was about to be replaced. Separately, the Alibaba-linked ROME agent reached out to an outside computer on its own and started mining cryptocurrency until security systems caught and stopped it.
Should you be worried?
The good news is most of these incidents took place in contained test setups, and Reuters saw no sign of a Chinese-powered agent breaking loose online or dodging a shutdown.
Still, experts say we shouldn’t brush this off. Redwood Research’s Alex Mallen noted that these examples aren’t very dangerous today, but as agents grow more capable, their mischief gets smarter and tougher for us to deal with.
AI agents seem to share the same bad habits no matter where they’re built, and as they take on more real-world tasks, keeping them honest could become a real challenge.
Related Coverage
- AI models keep escaping their sandboxes, and Kimi K3 is the latest to join the party
- DeepSeek V4 is out, touting some disruptive wins over Gemini, ChatGPT, and Claude
- OpenAI pauses advanced model development after AI agents bypass safeguards again
- OpenAI is working on a ‘kill switch’ after an AI escaped its test
