One-line summary
Per the China National Standard Information Public Service Platform, a batch of AI safety national standard plans was issued on September 30, 2026, including "Cybersecurity Technology - Artificial Intelligence Deep Synthesis Security Specification" (20264769-T-469), "Cybersecurity Technology - Artificial Intelligence Safety Guardrail Technical Requirements" (20264765-T-469), and "Cybersecurity Technology - Artificial Intelligence Training and Inference Framework Security Requirements" (20264762-T-469), all currently in drafting.
Standard overview
| Item | Content |
|---|---|
| Plan No. (1) | 20264769-T-469 |
| Plan name (1) | Cybersecurity Technology - Artificial Intelligence Deep Synthesis Security Specification |
| Plan No. (2) | 20264765-T-469 |
| Plan name (2) | Cybersecurity Technology - Artificial Intelligence Safety Guardrail Technical Requirements |
| Plan No. (3) | 20264762-T-469 |
| Plan name (3) | Cybersecurity Technology - Artificial Intelligence Training and Inference Framework Security Requirements |
| Plan No. (4) | 20264132-T-469 |
| Plan name (4) | Artificial Intelligence - Application Specification for Data Processing Units for Intelligent Computing Centers |
| Plan No. (5) | 20264903-T-469 |
| Plan name (5) | Intelligent Computing Cluster - High-Speed Network Management Technical Specification |
| Issuing body | State Administration for Market Regulation / National Standardization Administration (published on the national standards platform) |
| Type | New development (制定) |
| Plan issue date | September 30, 2026 |
| Status | Drafting (正在起草) |
Scope of application
- 20264769-T-469 (deep synthesis security): applies to internet information service providers and technical systems that offer deep synthesis services (face synthesis, voice synthesis, image/video generation, etc.).
- 20264765-T-469 (safety guardrails): applies to the design and evaluation of guardrails for LLM applications, covering content safety, refusal behavior and suppression of harmful outputs.
- 20264762-T-469 (training and inference framework security): applies to developers and deployers of AI training and inference frameworks, covering framework vulnerabilities, supply-chain security and runtime environment security.
- 20264132-T-469 and 20264903-T-469: apply to organizations planning, building or operating data processing units (DPUs) in intelligent computing centers and high-speed networks (e.g., RDMA cluster networks).
- Per the national standards platform search results, participating drafting organizations include Ant Group (20264769-T-469), Alibaba (20264765-T-469), Baidu (20264762-T-469) and Tianyi Cloud (20264132-T-469, 20264903-T-469), subject to the final published version.
Key clause points
| Clause | Content | Affected parties |
|---|---|---|
| 20264769-T-469 scope | Deep synthesis security specification: technical requirements for synthetic content labeling, authenticity verification and abuse prevention | Deep synthesis providers, audio/video synthesis platforms |
| 20264765-T-469 scope | AI safety guardrail technical requirements: security boundaries, interception policies and evaluation methods for LLM applications | LLM application developers, API providers |
| 20264762-T-469 scope | Training/inference framework security: supply-chain security, runtime protection and vulnerability management | Model training/inference platforms, framework community users |
| 20264132-T-469 scope | DPU application specification for intelligent computing centers: offload, acceleration and security capabilities | Intelligent computing center builders, cloud providers |
| 20264903-T-469 scope | High-speed network management for intelligent computing clusters: topology, performance and operations | Cluster operators, HPC/AI data centers |
Compliance recommendations
- Deep synthesis and generative AI providers: track drafts of 20264769-T-469 and 20264765-T-469; benchmark against the current Interim Measures for the Management of Generative AI Services, covering synthetic content labeling and content-safety interception.
- Model training/inference platforms: use 20264762-T-469 as a direction to inventory framework supply chains and vulnerability-response processes, preparing for compliance audits after the formal standard is published.
- Intelligent computing centers and cloud providers: watch 20264132-T-469 and 20264903-T-469; reserve standard compatibility in DPU selection and cluster network builds. These are draft-status plans; the final numbering and clauses follow the published version.
- Set up a standard-tracking mechanism: national standard development typically takes 12-24 months after issuance; check the national standards platform quarterly for exposure drafts and submit feedback during the comment period.
Official sources
- National Standard Information Public Service Platform: standard plan search (20264769-T-469 etc.) link
- National Standard Information Public Service Platform: standard plan search (Alibaba plans) link
- National Standard Information Public Service Platform: standard plan search (Baidu plans) link
- National Standard Information Public Service Platform: standard plan search (Tianyi Cloud plans) link