On 29 February 2024 the National Technical Committee on Cybersecurity (TC260) published the Basic Security Requirements for Generative AI Services (TC260-003). It is a technical document rather than a departmental rule, but it changes how filing materials are prepared: the security assessment, previously written in general terms, now has items that can be checked one by one.
Six groups of requirements
The document divides requirements into corpus source security, corpus content security, corpus annotation security, model security, safeguards and security assessment. The corpus provisions address lawful and diverse sourcing, filtering and IP risk, and the qualifications and rules applied by annotators. The model provisions address training and output security. Safeguards cover management systems, content labelling, complaint channels and monitoring.
For companies preparing a large-model filing, this means the assessment can no longer say only that “necessary measures have been taken”. It must explain what each measure covers, how often it operates and how it is recorded.
How it maps to filing materials
Filing review does not measure length. It tests whether the materials corroborate one another. A corpus provenance statement should map to licence documents; annotation rules should map to annotation records; safeguards should map to actual complaint handling and content disposition. TC260-003 supplies that requirement-to-evidence mapping.
The common reasons for rejection cluster here: an incomplete licence chain, annotation rules that do not match practice, a complaint channel that cannot be reached, and generated content without visible or embedded labels.
Three steps for companies
- Run a gap assessment. Mark each of the six groups as present, partial or missing, then decide whether to fix the document or the process.
- Make records routine. Annotation, filtering, complaints and disposition cannot be assembled only before assessment; the timeline will not match system logs.
- Schedule labelling separately. Labelling duties do not depend on filing progress, and the absence of a label is itself a violation.
For route analysis and filing preparation see algorithm filing; for corpus licensing see data compliance. Definitions are in the glossary.
This article is based on the Basic Security Requirements for Generative AI Services (TC260-003) published by TC260 on 29 February 2024, and on public information. It is for reference only and does not constitute legal advice.