AI Safety Governance Framework 3.0 released as China updates health AI risk provisions

China’s National Technical Committee 260 on Cybersecurity of SAC (the Cybersecurity Standardisation Technical Committee, or CSTC) released the Artificial Intelligence Safety Governance Framework 3.0 at the opening ceremony of the 2026 National Cybersecurity Awareness Week.

The framework builds on two earlier editions, released in 2024 and 2025, which the CSTC says received wide attention domestically and internationally. Framework 3.0 was compiled under the guidance of the Cyberspace Administration of China, with input from research institutions and industry bodies including the China Academy of Cyberspace Studies and the Cyberspace Administration’s Data and Technology Support Centre.

The stated aim is to implement the Global AI Governance Initiative, respond to new trends in AI development, and address emerging governance challenges. The document retains the “risk classification, technical response, comprehensive governance” structure of its predecessors, while updating the risk categories and adjusting technical and governance measures.

Healthcare features in the framework as one of several named priority sectors rather than as a standalone focus. Under measures to promote standardised application in key industries, the framework calls for sector-specific safety guidelines to be developed for healthcare alongside government affairs, finance, education and broadcasting, covering model selection, deployment, operation and decommissioning.

Healthcare is also named among the sectors slated for tailored regulatory sandbox rules, with admission criteria and testing intensity to be set according to project risk level.

The framework separately flags medical surgery as one of the physical-interaction settings where embodied AI carries “physical actuation” risk, citing hallucinated decision-making, perception errors, control-algorithm flaws and hardware failure as potential causes of harm.

On talent development, it names “intelligent healthcare” alongside autonomous driving, brain-inspired intelligence and brain-computer interfaces as a frontier field requiring a stronger AI safety talent pipeline.

The bulk of the framework addresses AI risk more broadly, spanning model and data risks, agentic AI, cybersecurity, information content, and social, environmental, cultural and ethical risks arising from AI’s wider deployment, alongside corresponding technical and governance countermeasures.