China drafts AI export rules, banning 14 techs and restricting 26

China's Ministry of Commerce and the Ministry of Science and Technology unveiled draft revisions to the Catalogue of Technologies Prohibited and Restricted from Export in mid-October, opening the text to public comment. The proposal touches 40 technologies across AI, biomedicine and photovoltaics — 14 slated for a flat export ban, 26 restricted to licensed transfers. Automatic licensing and other fast-lane measures are pencilled in for particular cases.

The timing is no accident. Washington and Beijing are still trading blows over advanced chips, equipment and AI know-how, and the catalogue is one of China's main levers for keeping its edge in large models, autonomous agents and computing tech. The draft draws its authority from the Foreign Trade Law: banned tech stays home, restricted tech needs a licence, everything else moves freely with a contract on file.

The nuance lives in the detail. Large models trained on core or important data — the kind that can sponge up industrial processes, production know-how and trade secrets — get closer scrutiny, while general-purpose models built on ordinary commercial data don't pick up extra restrictions just for being AI. For autonomous agents, the draft weighs capability against where the system runs, targeting high-performance agents aimed at critical infrastructure, national defense and core government functions for a ban, and multi-agent coordination or tool use for licensing. Officials floated reference markers — giant compute workloads, models beyond two trillion parameters, frontier-level benchmark scores — but framed them as assessment aids, not a single technical tripwire.

Security worries run through the whole document. It points to Anthropic's Claude Mythos Preview, which, according to the company, surfaced a software vulnerability that had stayed buried for 27 years despite not being built as a cybersecurity tool. That kind of reasoning power in the wrong hands is exactly what Beijing says it wants to fence off, along with cyberattack and information-confrontation tech, which would face the strictest treatment.

Open source gets a softer landing, on paper. Tech already public through open-source releases or academic papers would sit outside export management, and restricted tech headed for open-source release could ride automatic licensing to cut paperwork. The draft also introduces project-based processing and advance submissions so regulators can move faster on applications.

Crucially, the catalogue targets the underlying tech, not the overseas sale of products. Chinese firms could keep selling abroad through APIs, software-as-a-service, AI applications, industry solutions and bundled hardware-software products — so long as no controlled tech actually crosses the border. The list of what counts as a substantive transfer: restricted model weights, sensitive training data, source code, key algorithmic documentation, computing-system designs and the guts of highly sensitive autonomous agents.