China GB/T 45654-2025 Basic Security Requirements for Generative AI Services 5.1: Model training security: content security as a core optimisation metric, framework audits and backdoor detection

What does 5.1 require?

During training the provider must treat the security of model-generated content as one of the main indicators for judging the quality of generated results, for example by building and continually updating a security risk test question bank, optimising the model against it and retesting after each optimisation, update or upgrade, and by building a security annotation dataset meeting clause 4.3 and using it for safety fine-tuning. It must regularly audit the security of the development frameworks and code it uses, with particular attention to the security of open-source frameworks and known vulnerabilities, and fix what it finds; and it must regularly test the model for backdoors and deal promptly with any found, for example through fine-tuning or unlearning.