China Drafts Strict Export Controls Capping AI Models Over 2 Trillion Parameters and Cyberattack Tools.

China Drafts Strict Export Restrictions on Advanced AI Technologies, High-Parameter Models, and Cyberattack Capabilities

China Ministry of Commerce (MOFCOM) alongside the Ministry of Science and Technology (MOST) has released a joint draft regulation establishing tight export controls on advanced high-technology sectors. Artificial intelligence represents the primary focus of the mandate, introducing explicit export prohibitions and conditional export licensing across offensive cyber-AI tools, military-grade systems, and trillion-parameter models.

Key Export Prohibitions, Restricted AI Categories, and Industry Impact

• Prohibited AI Technologies (Outright Banned from Export):

  • High-Capability Cyberattack AI: Specialized AI models designed for offensive security operations capable of achieving autonomous breach success rates exceeding 60% against multi-layered defense architectures, over 85% against industry-standard protective software suites, or possessing verified capabilities to execute sandbox evasion and security bypasses.

  • Models Trained on Sensitive Internal Data: Generative AI architectures pre-trained or fine-tuned on non-public Chinese domestic datasets or critical state-level informational assets.

  • Defense and National Security AI: Tactical AI systems developed for direct military integration, autonomous warfare operations, or state surveillance infrastructure.

• Restricted AI Technologies (Subject to Mandatory Licensing Approval):

  • Intermediate Cyberattack AI: Systems capable of autonomous vulnerability discovery, multi-stage exploit chaining, or achieving over 75% success rates in penetration testing benchmarks.

  • Ultra-High Precision Voice Recognition: Audio processing and acoustic AI models achieving recognition accuracy rates equal to or exceeding 99.1% with error rates under 0.01%.

  • Frontier Multi-Trillion Parameter Models: High-capacity AI architectures utilizing strategic national training data, exceeding 2 trillion (2T) total parameters, or consuming total training compute exceeding 10²⁶ FLOPS.

  • Related High-Tech Frontiers: Additional non-AI restricted categories include advanced semiconductor manufacturing techniques, specialized telecommunication protocols, gene therapy platforms, and high-performance materials science.

• Implications for the Global Open-Weight AI Ecosystem:

  • Impact on Chinese Open-Weight Models: The regulatory 2-trillion parameter boundary poses immediate strategic questions for Chinese tech giants that historically released frontier open-weight models to the global developer community.

  • Capping Open Releases: While ultra-large foundation models like Kimi K3 (2.8T) cross the regulatory 2T threshold and face strict export licensing, high-performing open-weight architectures sitting under the cap such as DeepSeek-V4 Pro (1.6T) remain positioned to operate within permissive export boundaries.

 

Source: MOFCOM.gov.cn 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

China Drafts Strict Export Restrictions on Advanced AI Technologies, High-Parameter Models, and Cyberattack Capabilities

China Ministry of Commerce (MOFCOM) alongside the Ministry of Science and Technology (MOST) has released a joint draft regulation establishing tight export controls on advanced high-technology sectors. Artificial intelligence represents the primary focus of the mandate, introducing explicit export prohibitions and conditional export licensing across offensive cyber-AI tools, military-grade systems, and trillion-parameter models.

Key Export Prohibitions, Restricted AI Categories, and Industry Impact

• Prohibited AI Technologies (Outright Banned from Export):

  • High-Capability Cyberattack AI: Specialized AI models designed for offensive security operations capable of achieving autonomous breach success rates exceeding 60% against multi-layered defense architectures, over 85% against industry-standard protective software suites, or possessing verified capabilities to execute sandbox evasion and security bypasses.

  • Models Trained on Sensitive Internal Data: Generative AI architectures pre-trained or fine-tuned on non-public Chinese domestic datasets or critical state-level informational assets.

  • Defense and National Security AI: Tactical AI systems developed for direct military integration, autonomous warfare operations, or state surveillance infrastructure.

• Restricted AI Technologies (Subject to Mandatory Licensing Approval):

  • Intermediate Cyberattack AI: Systems capable of autonomous vulnerability discovery, multi-stage exploit chaining, or achieving over 75% success rates in penetration testing benchmarks.

  • Ultra-High Precision Voice Recognition: Audio processing and acoustic AI models achieving recognition accuracy rates equal to or exceeding 99.1% with error rates under 0.01%.

  • Frontier Multi-Trillion Parameter Models: High-capacity AI architectures utilizing strategic national training data, exceeding 2 trillion (2T) total parameters, or consuming total training compute exceeding 10²⁶ FLOPS.

  • Related High-Tech Frontiers: Additional non-AI restricted categories include advanced semiconductor manufacturing techniques, specialized telecommunication protocols, gene therapy platforms, and high-performance materials science.

• Implications for the Global Open-Weight AI Ecosystem:

  • Impact on Chinese Open-Weight Models: The regulatory 2-trillion parameter boundary poses immediate strategic questions for Chinese tech giants that historically released frontier open-weight models to the global developer community.

  • Capping Open Releases: While ultra-large foundation models like Kimi K3 (2.8T) cross the regulatory 2T threshold and face strict export licensing, high-performing open-weight architectures sitting under the cap such as DeepSeek-V4 Pro (1.6T) remain positioned to operate within permissive export boundaries.

 

Source: MOFCOM.gov.cn