AI Safety Governance Framework 2.0

The first framework was released in September 2024. Under Cyberspace Administration of China guidance, CNCERT led specialist institutes, research bodies, and companies in preparing version 2.0; the official bilingual PDF names TC260 and CNCERT/CC. The update was released on 15 September 2025. The official announcement says it refined risk classification, explored risk grading, and updated governance measures in response to technical and application changes. AI Safety Governance Framework 2.0 is the official English title, not a translation supplied by Digital Policy Alert.

Sources: s1, s2, s3, s4

The framework provides a structured view of priorities within China's AI policy and standards community. It groups inherent model, algorithm, and data risks; application risks involving cyber systems, content, the physical world, and cognition; and derivative social, environmental, and ethical risks. It connects those categories to technical and broader governance measures and emphasizes the full AI lifecycle. Independent analyses treat it as a possible precursor or reference point for later standards and rules. A standalone explanation prevents readers from mistaking policy guidance for an enforceable duty.

Sources: s3, s4, s6

An AI provider could use the framework voluntarily to structure a risk register: map a use case to the listed risk families, assess likelihood and impact, assign technical and organizational measures, and revisit them from research through operation. That exercise may support gap analysis, but it does not by itself establish conformity with Chinese law or any mandatory standard. An auditor making a compliance claim would need to identify the actually applicable statutes, administrative measures, standards, contracts, and facts separately; adoption of the framework is neither a legal safe harbor nor a safety certificate.

Sources: s2, s3, s4, s5

The official English text accompanies the Chinese release and is credible for terminology, but contested legal interpretation should still consult the Chinese original and subsequent instruments. The framework records recommended categories and measures; it does not show that any control is effective or that a system is safe. Later TC260 standards, administrative measures, or legislation may change its practical relevance. This entry reflects sources checked through 5 September 2026 and should not be used as legal advice, a jurisdiction-specific compliance opinion, or safety certification.

Sources: s2, s4, s5, s6