Security · November 19, 2026

Chinese AI models restrict cyberattack capabilities amid export control efforts

grey metal chain on brown wooden door
Sheldon Kennedy / Unsplash

Major Chinese artificial intelligence companies have limited cyberattack functions in their latest flagship models while strengthening vulnerability detection and defense capabilities, according to reports on November 17.

Alibaba's newest flagship model Qwen4-Max deliberately filtered training data containing material usable for cyberattacks while preserving abilities useful for vulnerability detection scenarios. The company has not publicly released model weights, preventing users from downloading and deploying unreviewed versions. Alibaba's Qwen series was previously among the world's most downloaded open-weight model families.

Tencent restricted network security functions in Hy4 before its official release and removed vulnerability exploitation generation data from its training process. Moonshot AI removed cyberattack-related materials from two stages of post-training when announcing Kimi K3.1 Pro on October 23. According to the company, the filtered model scored 81.0 percent on internal vulnerability exploitation benchmarks, lower than the unfiltered version's 90.5 percent.

Zhipu conducted security reviews before releasing GLM-5.3 weights in August and reported system vulnerabilities to relevant authorities.

CrowdStrike disclosed in an October report that an organization called Vapor Panda used DeepSeek models with the Artex framework to launch intrusions against more than 200 institutions across at least 14 countries within an 11-day window.

Tencent's chief artificial intelligence scientist stated on November 7 that models' cyberattack capabilities and forthcoming biological capabilities would approach thresholds that could cause actual harm.

Fangcun AI's founder told Bloomberg that removing attack materials from training represents a balance between cost and effectiveness, but no method currently exists to completely eliminate model risks.

In October, China's Ministry of Commerce and Ministry of Science and Technology announced plans to revise the prohibited and restricted technology export list, proposing to include artificial intelligence-driven malicious code generation and mutation technologies in export controls or restrictions, affecting 23 technologies in the artificial intelligence field. In early November, officials stated they were accelerating safety standards and ethical guidelines for artificial intelligence developers, emphasizing embedding security measures during development.