Cybersecurity technology — Basic security requirements for generative artificial intelligence service

Summary

Recommended national standard setting basic security requirements for generative AI services, aimed mainly at services with public-opinion attributes or social-mobilisation capacity and supporting filing management and testing and assessment. It covers training-data security (no sources with over 5% illegal and unhealthy information, intellectual property, personal information, annotation), model security (at least 90% of sampled outputs free of the listed risks, refusal of clearly extreme or inducing questions, monitoring, backdoor checks) and measures such as transparency, protection of minors and complaints. Annexes list 31 main security risks and assessment methods.

Language

English translation by Center for Security and Emerging Technology (CSET), Georgetown University, not by a public body Open it. The Chinese text is authoritative.

Safety topics

Status notes

Published 25 Apr 2025 and effective 1 Nov 2025; TC260 posted the full text on 30 Jun 2025. It is a recommended 'GB/T' national standard, not a mandatory 'GB' one, and its annexes on risks and assessment methods are informative. It was formulated on the basis of TC260's February 2024 technical document TC260-003 (CN-06). Its introduction names two companion standards: GB/T 45674 on generative AI data annotation security and GB/T 45652 on the security of pre-training and fine-tuning data.

Sources

Open the official source https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F67D3F376E0A0A0FF5317FB36B32A30A

Secondary source: https://www.tc260.org.cn/tc260/xwdt1/202506/739067714b9f40daa20219ea8fd7bf6d.shtml

Local copy

Verification

History

1. 2026-10-14related: [{"slug": "cn-tc260-003-genai-security-requirements", "id": "CN-06"}, {"slug": " → [{"slug": "cn-tc260-003-genai-security-requirements", "id": "CN-06"}, {"slug": " — two-way link to CN-18, CN-19, CN-20, CN-21, CN-22, CN-23, CN-24, CN-25 added in run 2026-10-14-discover-leads-2026-10-142026-10-14-discover-leads-2026-10-14 · Claude Code (claude-opus-5-5)
2. 2026-10-14english_version: – → {"kind": "third-party translation", "url": "https://cset.georgetown.edu/publicat — Return round (14 Oct 2026): set to CSET's unofficial translation, now stored (page and PDF; translator Etcetera Language Group, Inc.). The SAMR catalogue gives only the official English title.2026-10-14-refresh-translation-fixes · Claude Code (claude-opus-5-5)
3. 2026-10-14notes: Published 25 Apr 2025, effective 1 Nov 2025; TC260 posted the full text on 30 Ju → Published 25 Apr 2025 and effective 1 Nov 2025; TC260 posted the full text on 30 — Return round (14 Oct 2026): rewritten: added that the annexes are informative (资料性), the basis in TC260-003 (CSET translator's note) and the companion standards as the introduction names them. Removed 'used in regulatory filing practice', the companion standards' dates and 'both in force' (in no stored file), and the 'listed on SAMR platforms' source remark.2026-10-14-refresh-translation-fixes · Claude Code (claude-opus-5-5)
4. 2026-10-14safety_topics: ["general-purpose / frontier AI", "evaluation & testing"] → ["general-purpose / frontier AI", "evaluation & testing", "transparency & labell — Return round (14 Oct 2026): added 'transparency & labelling' (6.2 service transparency; 5.2 e labelling of generated images and video) and 'child safety' (6.1 c and d: anti-addiction settings for minors, keeping minors off unsuitable services).2026-10-14-refresh-translation-fixes · Claude Code (claude-opus-5-5)
5. 2026-10-14summary: Recommended national standard that sets baseline security requirements for gener → Recommended national standard setting basic security requirements for generative — Return round (14 Oct 2026): rewritten from the standard's Chinese text, now stored. Dropped 'turns TC260-003 into a national standard' (moved to notes, on CSET's translator's note) and 'main technical reference for compliance with the Generative AI Measures' (the introduction says it supports filing management and testing and assessment, 支撑备案管理、检测评估). Added the target services (引言), the 5% source threshold (4.1.1), the 90% pass rate (5.2 a), refusal (5.2 d), minors (6.1 c) and transparency (6.2). 安全 stays 'security', as in the official English title.2026-10-14-refresh-translation-fixes · Claude Code (claude-opus-5-5)
6. 2026-10-14issuing_body: State Administration for Market Regulation / Standardization Administration of C → State Administration for Market Regulation / Standardization Administration of C — Return round (14 Oct 2026): 'drafted by TC260' replaced by 'proposed and administered by TC260' (前言: 由全国网络安全标准化技术委员会(SAC/TC260)提出并归口); the drafting units are CESI and others.2026-10-14-refresh-translation-fixes · Claude Code (claude-opus-5-5)

Related records

TC260-003 China · Superseded

Generative AI Measures China · In force

Draft national standard on AI security capability maturity assessment China · Draft / proposed

TC260 draft guide on agentic AI system development security China · Draft / proposed

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

TC260-PG-20269A (AI application security, education) China · Published (non-binding, current)

TC260-PG-202610A (AI application security, health) China · Published (non-binding, current)

TC260-PG-202611A (AI application security, radio, TV and network audiovisual) China · Published (non-binding, current)

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

TC260-PG-202515A (generative AI security emergency response) China · Published (non-binding, current)

Citation

State Administration for Market Regulation / Standardization Administration of China (proposed and administered by TC260). Cybersecurity technology — Basic security requirements for generative artificial intelligence service [GB/T 45654-2025]. 2025-04-25. https://openstd.samr.gov.cn/bzgk/std/newGbInfo?hcno=F67D3F376E0A0A0FF5317FB36B32A30A. Status: Published (non-binding, current); verified 2026-10-14 (level A). Local copy archive/CN-10/2026-10-11/official.html retrieved 2026-10-11 14:00:13 UTC, SHA-256 be053ff096fd143f2a743e957f6c9dffe66228fbf2ca91890595f70a2fb820f0. Every AI Policy Ever, version 1.18 (snapshot 2026-10-18), record CN-10.

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.