China's NFRA issues first AI security guidelines for banks and insurers

China's National Financial Regulatory Administration (NFRA) issued its first dedicated AI security framework for the banking and insurance sectors, the Guiding Opinions on the Safe Development and Application of Artificial Intelligence in the Banking and Insurance Industries.

The document was released as regulatory filing Jin Fa [2026] No. 8.

What it is

The Guiding Opinions set out 32 measures organized under seven regulatory areas covering governance, development, data, computing, risk, capability building and supervision.

Although framed as guidance, legal analysts describe it as a binding compliance document for covered institutions.

Banking and insurance institutions developing or applying AI must follow stated principles for the technology's use.

What it requires

Institutions must strengthen top-level planning across the organization as part of overall AI governance.

For higher-risk uses, institutions must apply tiered risk controls and put humans in the loop at key decision points.

The guidelines give, for the first time, a defined list of high-risk scenarios that includes fund trading, credit approval, and underwriting and claims settlement.

AI applications in those areas must be approved by an institution's risk management committee and reported to the regulator.

On data, the rules bar use of sensitive personal information, such as names and identification numbers, to train or tune generative AI models.

Externally sourced generative AI models must also be registered with the Cyberspace Administration.

Institutions are encouraged to build their own self-reliant, controllable, secure and efficient computing power base as needed.

What security leaders should do

Security and risk teams at any bank or insurer operating in China, including branches of foreign firms, should map existing AI use cases against the newly defined high-risk categories and confirm which ones now need risk-committee sign-off and NFRA reporting. Data teams should audit training pipelines for any generative AI model to confirm that names, identification numbers or similar personal data are not being used, and that externally sourced models have been registered as required. Because the framework spans governance, computing infrastructure, data, and outsourcing, CISOs should treat it as a companion to existing NFRA data-security obligations rather than a standalone checklist, and build human-oversight checkpoints directly into AI deployment workflows for fund trading, credit, underwriting and claims processes.