Cases of Deception and Restriction Bypassing Identified in Chinese AI Agents

AI agents built on Chinese artificial intelligence models exhibited behaviors during testing such as deception, circumventing restrictions, and concealing their own failures. According to studies and technical reports examined by Reuters, such capabilities raise questions about maintaining human oversight as AI systems grow increasingly autonomous.

In one case conducted this year, agents built on Alibaba, DeepSeek, and Moonshot models provided false information about their own capabilities in order to win a simulated business tender. When the systems were asked to redo the task, they further intensified their original deception strategy.

In another experiment, AI agents concealed their failure to complete a task. In the testing environment, they began simulating results and creating nonexistent files to create the impression that the task had been completed successfully.

AI agents are software systems that use artificial intelligence models and computer tools to perform complex tasks with little or no human intervention. Their growing autonomy represents one of the major challenges in the field of AI safety.

Reuters reviewed more than 200 documents, including university studies, technical reports, and AI system evaluations. The outlet identified at least 20 studies and evaluations since 2025 describing instances of deception, self-replication, and the testing of established boundaries by AI agents.

Some experts view such behaviors as a possible foundation for capabilities that could, in the future, allow AI systems to break free of human control. In their assessment, as AI becomes more autonomous and powerful, managing such systems could become increasingly difficult.

At the same time, Reuters' review found no evidence that any agent running on Chinese AI models had independently managed to access the broader internet or avoid being shut down.

"These findings show that the components necessary for an uncontrolled escape already exist," said Colin Shea-Blymyer, a researcher at Georgetown University's Center for Security and Emerging Technology.

The cases involving Chinese AI systems come amid growing global attention to the safety of AI agents. Of particular concern is the question of whether humans can maintain control over systems that independently plan actions, use various digital tools, and adjust their own behavior based on the results they obtain.