China’s Ministry of Commerce (MOFCOM) and Ministry of Science and Technology have proposed controls on exports of AI cyber tools as Z.ai’s downloadable GLM-5.3 draws attention for exploit performance in U.S. evaluations.
The ministries opened a consultation Oct. 16 on adding 40 technologies spanning artificial intelligence, biomedicine and solar power to the export catalogue. Twenty-three entries are AI-related; across the three fields, 14 technologies would be prohibited and 26 restricted. The notice also requires advance registration for specified exports during the interim period before the revised catalogue takes effect.
MOFCOM identifies AI technologies for generating or altering malicious code as candidates for prohibition, while AI-assisted software vulnerability discovery would face proposed restrictions. The ministry says the technologies could be used to attack critical infrastructure or to bypass national security defenses. It also says the policy is intended to preserve normal international technology cooperation.
Under MOFCOM’s explanation, technologies already made public through open-source releases or academic papers would be outside the proposed export management. A restricted technology sent abroad for local deployment would be covered, the ministry says; remote use through an API, app, website or SaaS offering would not count as exporting the model technology.
MOFCOM says it has revised the export catalogue several times since 2001 and describes the current review as routine.
The National Institute of Standards and Technology (NIST) says Z.ai released GLM-5.3 on Aug. 14 and made its weights available for download two weeks later. NIST’s Center for AI Standards and Innovation found it was the leading open-weight model for cyber capability among those evaluated, but its combined benchmark results placed it about four months behind the U.S. frontier. Z.ai said the model’s cyber capabilities advanced faster than expected during post-training.
Anthropic recorded successful end-to-end exploitation in 50 of GLM-5.3’s 410 ExploitBench attempts and 56 of Claude Mythos Preview’s 410. In a separate 100-task test, full control-flow hijacks occurred in 4% of GLM-5.3 trials and 6% of Claude Mythos Preview trials. Anthropic said the evaluations ran in isolated, sandboxed environments, measuring test performance rather than confirming real-world misuse.
The proposal does not name GLM-5.3. Public comments are due Nov. 15, and the ministries have not issued the final catalogue.
