In September 2025 the National Technical Committee 260 on Cybersecurity (TC260) released a technical document, Security Specification for Government Large Model Applications. It sets out security requirements and test methods for government large model applications across four stages — model selection, application deployment, application operation and decommissioning — and applies to the design, development and testing of such applications.
Why the public sector needs its own baseline
Government use differs in three ways: data sensitivity is high, often involving unpublished administrative information and personal data; the service user is a public body, so the social impact of error is amplified; and procurement chains are long, involving builders, operators and vendors.
A general large model security standard answers whether the model itself is safe. The public sector also needs to answer whether it is safe to use in administrative processes. The technical document makes the latter concrete.
Requirements across four stages
Selection concerns whether a model meets the capability and compliance needs of government use, including assessment of model provenance, training data compliance and service stability.
Deployment concerns the security of the deployment environment, including configuration of compute and network environments and boundary management for data access.
Operation concerns continuous security, including output interception, log retention, anomaly monitoring and incident response. This echoes the requirements of the national generative AI security standard. See our note on the generative AI security standard.
Decommissioning concerns exit security, including data cleansing, model retirement and information protection during migration. This stage is the weakest in most corporate practice, yet it is critical for government data.
What it means for AI vendors
AI companies serving the public sector are usually vendors of models or systems rather than end users. In practical terms the document requires them to:
provide verifiable evidence of security capability rather than commitments in a bid document; cooperate with government security testing, for which the methods are now indicated; and maintain security management documentation across the full lifecycle, especially a decommissioning plan.
Government security requirements also stack with duties under the Cybersecurity Law and the Data Security Law, so companies should avoid the mismatch of “meeting the standard but not the legal duty”. See our notes on the AI clause in the revised Cybersecurity Law and network data security risk assessment.
A broader signal
A security specification for government large models is a prerequisite for “AI plus government” moving from pilots to scale. Technical documents of this kind typically precede mandatory requirements and function as advance guidance. See our notes on the State Council’s “AI plus” deployment and AI compliance in public governance.
For companies in government projects, building a self-check list across the four stages beats remedial work at the acceptance stage.
This article is based on TC260’s technical document Security Specification for Government Large Model Applications and on public information. It is for general reference only and does not constitute legal advice.