Safety framework covering AI agent risks necessary

Editor’s note: The Chinese government issued the Artificial Intelligence Safety Governance Framework 3.0 at the opening ceremony of the 2026 China Cybersecurity Week on Monday in Jinan, capital of East China’s Shandong province. The framework outlines principles for artificial intelligence governance, identifies potential risks and recommends measures to mitigate these risks. China Central Television spoke to Wen Yuheng, an associate professor of the AI law research institute at the China University of Political Science and Law, about what’s new in version 3.0 of the framework. Below are excerpts of the interview. The views don’t necessarily represent those of China Daily.

The release of three versions of the AI Safety Governance Framework in as many years reflects the country’s efforts to adapt its governance strategies to the rapidly evolving AI landscape.

Framework 1.0 released in 2024 focused on the inherent risks of both AI and the application of the technology.

Framework 2.0 expanded on this by addressing derivative risks, such as challenges to employment and ethical considerations, while establishing fundamental principles for trustworthy AI.

Since then, the world has witnessed major leaps in AI’s capabilities. Large models have moved beyond merely answering questions and entered a new phase where agents can autonomously perform tasks. Therefore, Framework 3.0 pays closer attention to the safety risks posed by AI agents.

Framework 3.0 clearly identifies the risks associated with agentic and embodied AI, signaling a structural shift in AI governance.

Previously, during the information age and the internet era, practices and information were projected from the real world onto virtual cyberspace. But now, as AI agents’ capabilities advance, actions generated by the agents in the virtual space have a tangible impact in the real world, introducing new risks.

Although Framework 3.0 is not mandatory, it serves as a benchmark. Regulators may follow this guideline when managing the risks of AI products and overseeing related business activities.

Also, the framework will help companies win the trust of users if they adopt the recommendations in the newly released guideline.

Framework 3.0 also calls for strengthening the management of open-source ecosystem safety and supply chain safety. When it comes to safety and innovation, the goal is not to find a fixed balance point. Instead, a dynamic approach is needed. The level of safety measures should be determined by the level of specific risks identified during the process of innovation, thus maintaining a balance between safety and innovation.

The emphasis on safety does not stifle innovation. Without ensuring safety, innovation could become a potentially destructive force.

To promote international cooperation on AI technology and governance, major countries need to build mutual trust. They should also learn from each other and draw on each other’s governance practices.