Researchers: Even Chinese AI agents can lie and circumvent rules

AI agents based on models from companies such as Chinese Alibaba, Deepseek, and Moonshot have reportedly demonstrated an ability to lie and circumvent their security limitations in controlled tests. A news agency has reviewed over 200 research reports and other documents, finding at least 20 studies or evaluations since 2025 where Chinese AI models exhibited behaviors such as fraud, replication, or attempts to bypass restrictions. However, there is no evidence that any Chinese AI agent has independently managed to escape onto the open internet or evade human control.

This is reminiscent of the incident where American OpenAI's AI agents broke out of their testing environment and intruded on the AI platform Hugging Face. China's latest framework for AI safety has identified AI agents that deceive evaluators, conceal their capabilities, or acquire resources and permissions independently as potential security issues.