On 1 November 2025 the national standard Cybersecurity Technology — Basic Security Requirements for Generative Artificial Intelligence Services (GB/T 45654—2025) took effect. It is one of the foundational national standards in the field of generative AI security.
What the standard changes
Before the standard, the security requirements for generative AI services came mainly from the Interim Measures for the Administration of Generative AI Services and similar rules, which are largely principle-based — “take effective measures to prevent the generation of unlawful content”, “improve the accuracy of generated content”. Principles have the advantage of broad coverage; their weakness is that a company can struggle to prove it has done enough.
GB/T 45654 breaks those principles into testable technical clauses covering corpus sourcing and security, generated-content safety, content labelling, user management and service transparency. It gives both companies and regulators a common technical yardstick for security assessment and inspection.
How it connects to existing rules
The standard is not isolated. It closes a loop with three existing regimes.
First, with the filing requirements under the Interim Measures. The safeguards a company must describe when filing can be implemented item by item against the standard — see our notes on filing and registration in practice.
Second, with the Measures for Labelling AI-Generated Synthetic Content. Labelling moves from “must do” to “how to do” — the technical paths for explicit and implicit labelling are now clearer. See our note on labelling enforcement.
Third, with the Cybersecurity Law and other foundational statutes. The revised Cybersecurity Law has brought AI within the foundational legal framework, and the standard extends it at the technical level — see our note on the revised Cybersecurity Law taking effect.
Four practical footholds
First, keep a traceable corpus ledger. The legality of training corpus sources is a focus of the standard. Companies should maintain a corpus register recording source channels, authorisation and cleaning rules, so that “unknown provenance” does not become a weakness in a regulatory inquiry.
Second, build interceptable content safety. The standard expects an ability to identify and handle unlawful or harmful content. That means retaining interception records and handling-process documentation, not merely declaring in the user terms that “generation of prohibited content is forbidden”.
Third, cover labelling across all output types. Text, image, audio and video outputs should all have labelling capability, avoiding a partial-labelling gap. See our note on the AI application ethics and security guideline.
Fourth, stand up assessment and governance. The standard sets more specific expectations for security assessment, so the governance structure needs a named owner, an assessment cycle and an archive mechanism.
Why act early
The standard is voluntary in nature and does not itself impose administrative penalties. But it will become an important reference in enforcement, judicial decision-making and industry procurement when assessing whether a company met its duty of care. In infringement disputes, meeting the standard often feeds directly into the finding of fault.
For generative AI companies, it is better to treat the standard as a self-check list — is the corpus evidenced, is content interceptable, is labelling complete, is assessment archived — than to explain after the fact. See our algorithm filing service and data compliance service.
The standard’s entry into force marks generative AI compliance moving into a “testable” phase. Compliance work shifts from writing documents to producing evidence.
This article is based on the national standard GB/T 45654—2025 and public information. It is for general reference only and does not constitute legal advice.