China proposes AI export controls and interim registration

On October 16, 2026, China's Ministry of Commerce and Ministry of Science and Technology issued Announcement No. 48 of 2026. The announcement proposes adding 40 technologies to China's catalogue of technologies prohibited or restricted from export, including 23 in artificial intelligence, and it imposes an interim pre-export registration requirement that is already in effect. The catalogue additions are open for comment until November 15, 2026, while the registration rule applies now.

What the draft would control

The draft covers 23 AI items alongside 5 biomedicine and 12 photovoltaic items, for 14 prohibited and 26 restricted items overall. Frontier model weights, checkpoints, source code, architecture, and training or inference optimization know-how would be restricted if a model meets any one of three tests: at least 10^26 FLOPs of training compute, at least 2 trillion parameters, or a global top-10 position on public evaluation rankings, global call volume, or comparable measures at the time of export. Agents are controlled separately. Planning, multi-agent coordination, tool use, and game-theoretic decision technologies would be restricted above set benchmarks, while high-performance agents for critical infrastructure, defense, or core government use would be prohibited. The annex also addresses AI-assisted vulnerability discovery, malicious code generation, detection technologies, deepfake detection, and chip or cluster technologies, and it excludes technology already in the public domain through open-source release or academic publication.

Service access versus custody

MOFCOM says API, app, web, SaaS, and lightweight on-device use do not count as exporting a large model. Local deployment of a restricted model outside China does count and would require a license. Automatic licensing may apply for deployment on a Chinese company's own overseas cloud or compute center, and for open-sourcing newly developed restricted technology after reporting. Weights already public are outside management scope. The dynamic top-10 test means a model's status can change after deployment, and commentators note that the rule names no ranking body.

Corporate and compliance exposure

Registration is triggered by transfers of ownership abroad, licensing for overseas use, and changes of ownership or control that make a foreign entity the owner or controller. Multinationals with China-based R&D teams may need to register before sharing code, checkpoints, or agent frameworks with overseas affiliates. The requirement also covers exports already in progress. Models trained on important data would be restricted unless a de minimis carve-out applies, while models trained on core data would be prohibited. In January 2026, MOFCOM assessed Meta's acquisition of Manus, a Singapore-registered agent company whose products were developed in China, for possible export authorization issues.

What organizations should do now

Organizations should inventory Chinese-origin models, agent frameworks, and AI security tools by consumption model and ask suppliers in writing whether they have registered exports. Legal and engineering teams should review technology flows out of China-based subsidiaries and R&D centers since October 16 for registration needs. Comments are due by November 15, 2026. Vendor contracts for locally deployed models should assign responsibility for registration and licensing and set notice rules if authorization is denied or revoked. Security teams should map China-developed detection and vulnerability-discovery tools against the thresholds and identify alternatives.

Sources: labs.cloudsecurityalliance.org