Key Takeaways
On October 16, 2026, China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology (MOST) issued Announcement No. 48 of 2026. The announcement proposes adding 40 technologies to the Catalogue of Technologies Prohibited or Restricted from Export, 23 of them in artificial intelligence, and it imposes an interim pre-export registration requirement that is already in effect [1][2]. China’s catalogue has touched AI before: the 2020 revision added personalized recommendation and AI interaction interface technology [6]. Unlike those 2020 additions, the draft reaches frontier model weights, agent architectures, and AI-enabled offensive cyber capability, with explicit technical thresholds for each [1].
The announcement contains two instruments with different legal status. The catalogue additions are a draft open for comment until November 15, 2026, while the registration requirement applies now, including to exports that have started but are not yet complete [1][2]. Within the draft, model weights are restricted at frontier scale. Weights, key checkpoints, source code, architecture, and training and inference optimization know-how are 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 [1][3]. Agents are controlled independently of the model. Planning, multi-agent coordination, tool use, and game-theoretic decision technologies are restricted above set benchmarks, and high-performance agents built for critical infrastructure, defense, or core government use are prohibited outright [1].
The line between service access and custody determines most enterprise exposure. MOFCOM says API, app, web, SaaS, and lightweight on-device use do not count as exporting a large model. Deploying a restricted model locally outside China does count, though automatic licensing may apply on a Chinese company’s own overseas cloud [2]. Multinationals also face exposure through licensing deals and changes of ownership or control, not only through product exports, and analysts read the registration form as potentially capturing intra-group code sharing as well [2][3][4]. Any transfer of ownership, licensing for overseas use, or shift of control to a foreign entity can trigger registration [2][4].
Background
China regulates outbound technology transfer under two overlapping regimes. The first is the technology-trade regime under the Foreign Trade Law and the Regulations on Technology Import and Export Administration, which governs the catalogue being revised here. The second is the dual-use export-control regime under the 2020 Export Control Law [5]. MOFCOM’s spokesperson described the revision as routine and grounded in Article 20 of the Foreign Trade Law. The spokesperson stressed that it “does not target any specific country, region, or entity” and that the aim is to prevent uncontrolled diffusion while protecting legitimate cooperation [2]. The catalogue has been revised several times since 2001. The August 2020 revision, which added recommendation-algorithm technology, coincided with the proposed sale of TikTok to a U.S. buyer and was reported as likely to complicate that deal [3][6].
Across all three sectors, Announcement No. 48 proposes 14 prohibited and 26 restricted items: 23 in AI, 5 in biomedicine, and 12 in photovoltaics [2]. In China’s technology-trade framework, prohibited technologies may not be exported at all, while restricted technologies require a license from commerce and science authorities. ChinaTalk notes that licensing is administered by provincial authorities, although central authorities have intervened in prominent cases [3]. The official annex groups the AI items into large models, agents, compute clusters, chip design, and information-processing technologies. It also excludes from management any technology that has already entered the public domain through open-source release or academic publication [1].
The interim registration requirement is the part that applies today. From the announcement date until the revised catalogue takes effect, exporters must register through MOFCOM’s unified business platform before exporting any listed technology. Registration covers four triggers: transfer of technology ownership abroad, licensing of domestic technology for overseas use, a change in ownership or control that makes a foreign entity the owner or controller, and other cases MOFCOM designates [2][4]. Exporters must decide for themselves whether a technology falls within scope. The registration form asks for recipient details, destination, technology description, end use, and estimated value [2][4].
The timing also matters for enterprises that already track U.S. controls. CSA’s prior research documented the first U.S. use of export-control authority to suspend a broadly deployed commercial frontier model [9]. Announcement No. 48 means frontier AI supply chains now face formal outbound controls from both of the largest AI-producing jurisdictions. ChinaTalk observes that the 10^26 FLOP threshold mirrors the figure used in the 2023 U.S. executive order and the since-rescinded 2025 U.S. AI Diffusion rule. The authors also estimate that no released Chinese model currently exceeds it [3].
Security Analysis
What the AI Annex Actually Controls
The annex is granular, and its structure matters for scoping. The table below summarizes the AI items drawn from the official draft annex [1], with interpretive notes from published analyses [3][4].
Category
Control
Key Thresholds or Conditions (draft)
Large models trained on China’s “core data”
Prohibited
Core data scope defined by the Data Security Law and related regulations
Large models trained on China’s “important data”
Restricted
Excludes trace use (≤0.1%) of desensitized important data that does not materially improve capability or reappear in outputs, and lawfully public data
Core large-model technologies (weights, checkpoints, source code, architecture, training and inference optimization)
Restricted
Any one of: ≥10^26 FLOPs; ≥2 trillion parameters; global top-10 position on public evaluation rankings, global call volume, or comparable measures at the time of export
High-performance agents for critical infrastructure, defense, or core government
Prohibited
Agent meets the restricted agent thresholds and is specially or mainly designed for those uses
Agent planning and task decomposition
Restricted
≥20-step subtask chains, ≥80% decomposition accuracy, ≥90% subtask success
Multi-agent collaboration architectures
Restricted
≥60 concurrent agents, completion time ≤2x the single-agent minimum, ≥85% agreement with reference answers
Tool calling and system operation
Restricted
≥200 integrated third-party tools, ≥90% tool-selection accuracy, ≥90% end-to-end success on tasks requiring 3+ tools
AI malicious code generation, adaptive evasion, and variant generation
Prohibited
Attack-success and automation thresholds, where usable against high-sensitivity systems
AI-assisted vulnerability discovery
Restricted
Any of six tests, including ≥75% on authoritative benchmarks, 10x senior-expert efficiency, or autonomous chaining of 2+ low or medium flaws into a working exploit
AI detection technology specially designed for cross-border data security
Prohibited
Detection accuracy and recall ≥95%
Deepfake detection; steganography; voiceprint; Chinese and minority-language translation
Restricted
Accuracy, recall, latency, and fidelity thresholds (for example, ≥98% for deepfake detection)
AI chip and memory chip design; large compute cluster technologies
Restricted
Process node ≤14 nm and other alternative criteria for AI chips; interconnect ≥800 GB/s, fault tolerance, and heterogeneous scheduling for clusters above 10,000 accelerators
Three features of this structure deserve attention from security and compliance teams. First, the “global leading” test is dynamic. A model that falls below the compute and parameter thresholds can still become restricted if its public evaluation rank, global call volume, or similar measures place it in the top 10 when the export occurs [1]. ChinaTalk calls this criterion “by far the most up-for-interpretation,” noting that the rule names no ranking body [3], and it means the compliance status of a given model version can change after deployment. Second, the agent items are defined by system behavior rather than by the underlying model. An orchestration framework that integrates hundreds of tools could meet the tool-use threshold even when it runs on a model well below the frontier thresholds, an issue Ray Sun identifies as unresolved [5]. Third, the cyber items carry direct security consequences: Chinese-origin AI red-teaming, vulnerability-discovery, and detection tools that meet the thresholds would require a license to leave China, and offensive code-generation capability could not be exported at all.
Deployment Models and Where the Line Falls
MOFCOM’s spokesperson framed the central test as whether restricted technology moves from inside China to overseas [2]. That framing produces a clear distinction between consuming a model as a service and taking custody of it, as summarized below.
Deployment scenario
Treatment under MOFCOM guidance
API calls, apps or websites, SaaS for overseas customers
Not within large-model export management [2]
Lightweight on-device deployment
Not within large-model export management [2]
Local deployment of a restricted model on infrastructure outside China
Within export management; license required [2]
Deployment on a Chinese company’s self-operated overseas cloud or compute center
Automatic licensing (“report and export”) may apply [2]
Open-sourcing newly developed restricted technology
Automatic licensing after reporting [2]
Technology already public via open-source release or academic papers
Outside management scope [1]
For enterprises, the practical consequence is that the deployment model the draft captures, local self-hosting, is also the one often chosen for data-protection reasons. Organizations that chose to self-host a Chinese frontier model in their own environment, rather than call a Chinese-hosted API, did so to keep sensitive data inside their own boundary. MOFCOM places local deployment outside China within export management, which means those self-hosting arrangements require the Chinese supplier to hold a license [2]. ChinaTalk reads this as also covering on-premises licensing deals and managed deployments in which a vendor runs an agent’s reasoning loop inside a customer environment [3]. Because licenses can be revoked or conditioned, ChinaTalk suggests this may lead some non-Chinese users to reconsider such dependencies [3]. The registration requirement already covers exports in progress, so ongoing weight-delivery or model-update arrangements may be affected now rather than after the catalogue is finalized [2].
The treatment of open weights depends on whether a model is already public. Weights that are already public are outside the controls [1]. Open-sourcing a new restricted model, however, falls under automatic licensing that still requires reporting [2]. Commentators note that “open source” is not defined, and that it is unclear how restricted-use licenses would be treated [3]. Organizations that depend on successive open-weight releases from Chinese labs should therefore treat future release cadence, not just current availability, as a variable outside their control.
Exposure Through Corporate Structure
The exposure most likely to surprise multinationals sits in their own organizational charts. Registration is triggered by licensing technology for overseas use and by changes of ownership or control, not only by sales to third parties [2][4]. A multinational with China-based R&D teams building agent frameworks, vulnerability-discovery
Texto de labs.cloudsecurityalliance.org tal como lo publicó, leído de su página y mostrado aquí con el formato de El Autómata. El original, en labs.cloudsecurityalliance.org · Volver a la noticia de El Autómata