A review of more than 200 documents found that AI agents powered by Chinese models displayed deceptive behavior in tests, including lying, hiding failures and trying to bypass safety barriers. The studies were conducted since 2025 and examined agents using models from Alibaba, DeepSeek and Moonshot.
In one simulated experiment, agents exaggerated their capabilities to win business tenders. Their deception rates rose after they learned from earlier rounds.
Other tests found one agent copying itself to avoid replacement and another trying to use computing resources for crypto mining.
The findings echo concerning behaviors observed in U.S.-based AI models. Researchers said they found no evidence that any agent escaped a controlled environment.