State of AI Safety in China (2026)

⚡ Quick Summary

Published by Concordia AI in July 2026, the fourth edition of the State of AI Safety in China tracks China’s evolving governance and technical landscape from July 2025 to June 2026. The report highlights a decisive pivot in Chinese AI governance: moving beyond content moderation and watermarking toward regulating autonomous action and agentic systems. Triggered in part by the rapid proliferation of open-source frameworks like OpenClaw, Chinese regulatory and standardization bodies have prioritized operational boundaries, intervention controls, and lifecycle security for autonomous agents. At the same time, top-level national strategy under the 15th Five-Year Plan balances economic diffusion under 'AI Plus' with full-lifecycle risk management, while international initiatives position Beijing as an active architect of multilateral AI governance frameworks.

🧩 What's Covered

The report provides detailed empirical data and policy analysis structured across five core domains:

Domestic Governance: Evaluates high-level mandates within the 15th Five-Year Plan (2026–2030), the introduction of dedicated Cyberspace Administration of China (CAC) guidance on agentic AI, binding administrative rules on anthropomorphic AI interaction services targeting emotional dependency and minor protection, mandatory pre-R&D ethics reviews across 10 pilot provinces, and proposed obligations in the draft Cybercrime Law for monitoring bulk malicious code generation.
National Standardization: Documents key moves by TC260 and SAC/TC28/SC42, including the formation of TC260’s dedicated AI Safety Working Group (WG9), the expansion of the AI Safety Governance Framework to Version 2.0 (introducing 'derivative risks' and circuit breakers), and forthcoming mandatory standards for agent application security and agent identity codes.
International Governance: Tracks China’s multilateral engagement at the United Nations—supporting the Independent International Scientific Panel on AI (IISP-AI) and proposing a World AI Cooperation Organization (WAICO)—alongside resumed bilateral dialogues with the United States and active participation in ISO/IEC and ITU-T technical standards.
Technical Safety Research: Analyzes over 900 arXiv papers from Chinese institutions, showing a 60% growth in monthly safety output. Research on autonomous agent safety surged from 8% to roughly 27% of quarterly output, accompanied by steady work on watermarking, mechanistic interpretability, and emerging evaluations for chemical, biological, radiological, and nuclear (CBRN) risks.
Industry Governance: Reviews collective self-regulation via the AI Industry Alliance of China (AIIA) and CAICT’s AI Safety Benchmark 2.0, while identifying persistent gaps and inconsistency in public model safety card disclosures among leading Chinese foundation model developers.

💡 Why it matters?

As Chinese open-source foundation models achieve global scale and lead worldwide download rankings, China's domestic safety benchmarks, technical research trajectories, and regulatory requirements directly influence global AI safety dynamics. The institutional shift from content moderation to 'action control' offers a concrete preview of how major jurisdictions manage autonomous agents, cybersecurity exposure, and dual-use biosecurity risks. For international policymakers, compliance directors, and technical safety teams, this report clarifies the structural realities, institutional actors (such as CAC, MIIT, CAICT, and TC260), and voluntary versus binding mechanisms operating inside China’s AI ecosystem.

❓ What's Missing

Because the report relies strictly on publicly documented policies, published academic literature, and official announcements, it cannot capture closed-door regulatory assessments or private pre-deployment evaluations conducted via China's mandatory algorithm registry. Furthermore, while high-level policy texts and national standards frameworks increasingly reference catastrophic and frontier risks (such as CBRN misuse or loss of control), detailed technical specifications and concrete enforcement mechanisms for these extreme risk tiers remain in early stages of development.

👥 Best For

AI governance officers, international policy analysts, regulatory affairs specialists, AI safety researchers, and risk managers overseeing global model deployment and cross-border AI compliance.

📄 Source Details

Authored by Gabriel Wagner, Erik Lindblad, Kwan Yee Ng, and Brian Tse; published by Concordia AI in July 2026. Interactive resources, databases, and policy matrices are available via the companion platform at aisafetychina.com.