Deals & regulation
China proposes export controls covering frontier AI models, agents and security tools
China has proposed adding 23 AI technologies to its restricted or prohibited export catalogue while immediately requiring registration for related transfers. The draft targets frontier model weights, training and inference know-how, agent architectures, and certain AI-enabled cyber tools, using compute, parameter, ranking, and capability thresholds. API access generally remains outside the proposal, but local deployment, overseas licensing, ownership changes, and some intra-group transfers may be covered. The catalogue is open for comment, and several boundaries—including “global leading” status and open-source treatment—remain uncertain.
Editorial summary · Publisher article
What matters
- The draft catalogue includes 14 prohibited or 26 restricted technologies across AI, biomedicine, and photovoltaics.
- Registration already applies to certain exports begun after October 16, including ownership transfers, overseas licensing, and control changes.
- Service access through APIs, apps, websites, and SaaS is generally distinguished from transferring model custody abroad.
- The proposal could affect multinational R&D sharing, self-hosted Chinese models, agent platforms, and AI security tools.
The main compliance distinction
The proposed framework generally leaves model use as a service outside export treatment but places locally deployed weights and code within the controlled boundary. That distinction could complicate privacy-motivated self-hosting arrangements and future model updates.
Why the scope may change
The catalogue remains a draft open for comments until November 15. Analysts also question how the dynamic top-10 test, open-source exclusions, agent thresholds, and intra-group code sharing will be interpreted in practice.
Context & caveats
- The catalogue additions are proposed rather than final, and some interpretations in the packet come from analysts rather than official guidance.
Why this made the edition
The packet provides extensive detail from the official draft and attributed outside analyses, distinguishes immediately effective registration from proposed catalogue changes, and explains practical effects for model custody, agents, corporate structures, and data. Inferences are explicitly identified.
Source material