Briefing #31: What China's new AI export controls actually mean (and it is not the end of open-source weights)

China's Ministry of Commerce and Ministry of Science and Technology released a draft revision of the Catalogue of Technologies Prohibited or Restricted from Export on October 16, proposing to add 40 entries, 23 of which are AI-related, covering some large language models, agents, chip design, compute clusters and offensive cyber tools. The draft has two tiers: 6 J-code entries are prohibited from export, and 17 X-code entries are restricted from export and require a license. Technologies already in the public domain through open source are not covered, and open-sourcing of restricted-category technologies is subject to automatic "report-and-export" licensing, so this draft is not the end of Chinese open-source models. The article assesses that some Chinese frontier models and agents are likely to fall into the restricted category, and whether they constitute the prohibited category is not yet clear and will depend on enforcement after the rules take effect. The article also mentions that Moonshot AI filtered cyberattack-related data from the training data of Kimi K3.1 Pro, and its self-built vulnerability discovery and exploitation benchmark score fell from 90.5% to 81.0%.