China’s National Information Security Standardization Technical Committee released the AI Safety Governance Framework 3.0 on 14 September 2026, updating guidance on AI risk management under the Cyberspace Administration of China’s supervision. Cyberspace Administration of China
The framework revises versions from 2024 and 2025 to address “new trends” in AI development and “new challenges” for safety governance, the committee said during the opening ceremony of the 2026 National Cybersecurity Awareness Week. Cyberspace Administration of China The text follows a “risk classification, technical response, comprehensive governance” structure and aims to strengthen “risk awareness” and ensure AI systems remain “safe and controllable.” Cyberspace Administration of China
Industry analysis by SemiAnalysis noted that the framework was drafted under the guidance of the Cyberspace Administration of China and is presented as part of implementing China’s Global AI Governance Initiative. Cyberspace Administration of China The commentary reported that the framework warns of “autonomous learning and recursive self-improvement” in AI systems and introduces a model-risk category for “behavior deviating from expectations,” including models that bypass protections or conceal capabilities. SemiAnalysis
The analysis argued that China’s regulatory posture puts “promoting AI innovation and development as the first priority,” citing the framework’s language and noting that comprehensive AI legislation listed in earlier legislative plans was not enacted in 2025. SemiAnalysis According to the firm, sectoral rules since 2025 have focused on content controls and application-level obligations rather than duties triggered by training compute or model capabilities. SemiAnalysis
SemiAnalysis reported building a dataset of 857 model releases from nine leading Chinese AI developers between 2021 and 15 September 2026 and finding that 31 releases, or 3.6%, were accompanied by a published safety evaluation result tied to that model. SemiAnalysis Nine releases, or 1.1% of the total, had such results available at or before launch, while 813 releases had no public safety disclosure under its criteria. SemiAnalysis The firm said that no Chinese frontier text model in its sample shipped with a published dangerous-capability evaluation in domains referenced by recent international AI safety statements. SemiAnalysis The official Framework 3.0 announcement does not address disclosure practices by individual labs or specify enforcement mechanisms or penalties for non-compliance. Cyberspace Administration of China