China released AI Safety Governance Framework 3.0 on 14 September 2026. It warns of self-accelerating AI and new risks like models bypassing controls. The framework still prioritizes rapid AI development as its first principle. China has issued many rules on AI content and applications since 2025. None of these rules address frontier model risks tied to compute or capability. A SemiAnalysis study found only 3.6% of 857 model releases from 9 Chinese labs had any safety result. Just 1.1% had results available at launch. No Chinese frontier text model has shipped with a dangerous-capability evaluation. Zhipu is the only lab with consistent safety statements and partial disclosures. Other labs rarely discuss frontier safety publicly.
What changed
China had no explicit frontier AI risk framework or systematic safety disclosures from labs.
What it unlocks
Understanding China's actual AI safety posture versus its stated policies and global comparisons.
- 857 model releases from 9 Chinese labs
- 3.6% of releases had any safety result
- 1.1% of releases had safety results at launch
- 70% agent penetration target by 2027
Sources