Every AI Policy Ever · open in the interactive register

Draft / proposedProposed · Technical standard

Cybersecurity technology — Assessment method for security capability maturity of artificial intelligence (draft for comment)

网络安全技术 人工智能安全能力成熟度评估方法

Record CN-18

Issued by
National Cybersecurity Standardization Technical Committee (TC260)
Jurisdiction
China
Level · region
National · East Asia
Published / adopted
2026-08-26
Language
Chinese

Summary

Draft recommended national standard, proposed by TC260, that sets out an AI security capability maturity model (ASMM) for assessing organisations, also usable by regulators as a reference. It covers data, model and application security in 15 security activities and 39 process areas, such as red-blue adversarial testing, value benchmark testing, backdoor detection, and security operations and emergency intervention. Each process area is assessed on the relevant ones of ten capabilities, among them robustness, explainability, fairness, rights protection and emerging-risk prevention, at five levels from informal execution to continuous optimisation, by self-assessment or a third party.

Language

No English version of this instrument was found. The English title and the summary are the register's own, written from the sources below; for the exact wording use the Chinese text.

Safety topics

evaluation & testinggeneral-purpose / frontier AIcatastrophic risk

Status notes

TC260 opened the draft for comment on 26 Aug 2026, with comments due by 25 Oct 2026. The draft has no standard number yet (GB/T XXXXX—XXXX) and leaves its publication and implementation dates blank. Organisations are typed as data-support, model-technology or application-service bodies, and each process area is mandatory, conditionally mandatory or optional for each type; an organisation reaches a level when all its applicable process areas reach it and no unaccepted major security risk remains. At the top level of post-training security, safeguards such as blocking knowledge of biological weapons and cyberattacks must not be removed or significantly weakened when the model runs in agents; at the top level of red-blue adversarial testing, organisations run human uplift trials that measure how far a model raises users' ability to write complex attack code or plan biological harm, and keep the uplift within safety thresholds. Normative references include GB/T 45654-2025 (CN-10), GB 45438-2025 (CN-08) and GB/T 45674-2025 (CN-24).

Sources

Open the official source https://www.tc260.org.cn/tc260-webapp/repo/file/TaskFile/0/20260828/dc90d8226fc547f3b23fe1b5c907154a.pdf

Local copy

Archived. Last attempt 2026-10-18 00:44:56 UTC

Official source (third-party copy, not published) archive/CN-18/2026-10-18/official.pdf
Retrieved 2026-10-18 00:12:47 UTC
SHA-256 9501da766283fee538403d87e63f47b11de3362796b629c3b81212d19c77e8fc

Verification

A
The instrument's own official text, or an official register, was stored and confirms the title, date and status. checked 2026-10-18 · official-url-fetched
Recheck of 18 Oct 2026. Stored and read: the draft for comment (tc260.org.cn, PDF of 76 pages, stored again on 18 Oct 2026 with the same SHA-256 as the copy of 14 Oct 2026), which gives the Chinese and English titles, the draft status (征求意见稿), the proposer and responsible committee (SAC/TC260), the blank number and dates, and the content behind the summary and notes; and TC260's JSON record of the consultation on project 70 (tc260.org.cn, stored in this run on 18 Oct 2026), which gives the comment period from 26 Aug to 25 Oct 2026, the title of the consultation notice and the draft PDF as the attached draft text. date_published is the start of the comment period. The explanatory note (.docx) has no text layer and is not used. The draft gives an English title only; no English text was found.
Second check: Independent check by Claude Code (claude-sonnet-5-5), 2026-10-18: 6 field(s) supported.

History

  1. 2026-10-18notes: TC260 opened the draft for comment on 26 Aug 2026, with comments due by 25 Oct 2 → TC260 opened the draft for comment on 26 Aug 2026, with comments due by 25 Oct 2 — the organisation-level rule completed with clause 9.6.2 c); the human uplift trials at the top level of PA17 added; GB/T 45674-2025 linked as CN-242026-10-18-recheck-east-asia-1 · Claude Code (claude-opus-5-5)
  2. 2026-10-18related: [{"id": "CN-10", "slug": "cn-gb-t-45654-2025"}, {"id": "CN-08", "slug": "cn-gb-4 → [{"id": "CN-10", "slug": "cn-gb-t-45654-2025"}, {"id": "CN-08", "slug": "cn-gb-4 — added CN-24 (GB/T 45674-2025, a normative reference) and CN-20 (TC260's AI application security guide, which refers to the national standard on AI security capability maturity assessment for maturity levels)2026-10-18-recheck-east-asia-1 · Claude Code (claude-opus-5-5)
  3. 2026-10-18safety_topics: ["evaluation & testing", "general-purpose / frontier AI"] → ["evaluation & testing", "general-purpose / frontier AI", "catastrophic risk"] — added catastrophic risk: the top levels require that safeguards against biological-weapon and cyberattack knowledge survive agent use, and human uplift trials on attack code and biological harm (PA15 and PA17, level 5)2026-10-18-recheck-east-asia-1 · Claude Code (claude-opus-5-5)
  4. 2026-10-18summary: Draft recommended national standard, proposed by TC260, that sets an AI security → Draft recommended national standard, proposed by TC260, that sets out an AI secu — corrected: the draft does not rate every process area on all ten capabilities; each selects the relevant ones (5.2.1 note); PA37 given its full name, 'security operations and emergency intervention'2026-10-18-recheck-east-asia-1 · Claude Code (claude-opus-5-5)
  5. 2026-10-14Added to the register — found in pipeline run 2026-10-14-discover-leads-2026-10-14, candidate C001 (CN/ministries)2026-10-14-discover-leads-2026-10-14 · Claude Code (claude-opus-5-5)

Related records

GB/T 45654-2025 China · Published (non-binding, current)

GB 45438-2025 (AI content labeling standard) China · In force

GB/T 45674-2025 China · Published (non-binding, current)

TC260-PG-20268A (AI application security, general provisions) China · Published (non-binding, current)

Citation

National Cybersecurity Standardization Technical Committee (TC260). Cybersecurity technology — Assessment method for security capability maturity of artificial intelligence (draft for comment). 2026-08-26. https://www.tc260.org.cn/tc260-webapp/repo/file/TaskFile/0/20260828/dc90d8226fc547f3b23fe1b5c907154a.pdf. Status: Draft / proposed; verified 2026-10-18 (level A). Local copy archive/CN-18/2026-10-18/official.pdf retrieved 2026-10-18 00:12:47 UTC, SHA-256 9501da766283fee538403d87e63f47b11de3362796b629c3b81212d19c77e8fc. Every AI Policy Ever, version 1.18 (snapshot 2026-10-18), record CN-18.

Improve this record

Know of an amendment, a status change or an error? Report news or an error on GitHub; the form opens with this record filled in.