China

China GB/T 45654-2025 Basic Security Requirements for Generative AI Services

25 controls. 0 other frameworks in our corpus share controls with it. Here is all of it, and how much of it you are already doing.

Page built . This page is derived from the framework corpus, which changes when the corpus is extended rather than daily.

25 controls 0 frameworks share controls with it China verified against its source document

Every control below is one this framework asks for. The right hand column counts how many other frameworks in our corpus carry the same control, which is the difference between doing this work once and doing it again for the next standard.

There is no implementation kit for this framework yet. The control list and the overlap above are free and complete.

What you already have

Frameworks whose controls overlap this one, most first. If you run any of them, the count is roughly what you have already evidenced.

No measured overlap with another framework in the corpus.

Every control

CodeControlAlso in
1-3Scope, references and terms0
4.1.1Training data source selection: sample before and after collection, reject sources above five per cent illegal or unhealthy content0
4.1.2Diversity and mix of training data sources0
4.1.3Management and traceability of training data sources: licences, collection records, commercial contracts and user authorisation0
4.2.1Filter all training data of every type before training0
4.2.2Intellectual property protection in training data and generated content0
4.2.3Personal information in training data only with consent or another lawful basis0
4.3.1Annotation personnel: training, assessment, qualification and separation of duties0
4.3.2Annotation rules for functional and security annotation0
4.3.3Accuracy of annotated content: batch sampling, re-annotation and review0
4.3.4Isolated storage of security annotation data0
5.1Model training security: content security as a core optimisation metric, framework audits and backdoor detection0
5.2Model output security: ninety per cent pass rate, accuracy, reliability, refusal behaviour and content labelling0
5.3Model monitoring and evaluation: input monitoring, routine testing and emergency handling0
5.4Model update and upgrade security: policy and self-assessment after major changes0
5.5Isolation of training and inference environments0
6.1Service suitability: justify use cases, protect high-risk scenarios and protect minors0
6.2Service transparency: publish audience, scenarios, purposes, limitations, model summary and personal information use0
6.3User input used for training only with a prominent notice and an easy opt-out0
6.4Public and user complaint and reporting channels with rules and time limits0
6.5Serving users: input detection with published suspension rules, and adequate monitoring staff0
6.6Service stability and continuity: backup and recovery0
6.7On-device models: activation, on-device security module, keyword updates and model patching0
AAppendix A: the 31 security risks in training data and generated content0
BAppendix B: security assessment reference method0

Tell me when China GB/T 45654-2025 Basic Security Requirements for Generative AI Services files something new

One email when a public company newly discloses something this framework governs, naming the company and what our corpus says it puts in scope. Nothing else, and one click to stop.

What this page is

A control-level reference for China GB/T 45654-2025 Basic Security Requirements for Generative AI Services, drawn from our framework corpus. Control codes and titles are references to the standard, not reproductions of it. The overlap counts and the auditor artefacts are our own work and are the part you will not find elsewhere.

Measure this against what you already run · All frameworks · Today's edition