From Startup Fortune · · 1m
Chinese AI Models Cut Hacking Skills
Chinese AI firms reduce cyber risks in new models for safer open weights.
In brief
Chinese AI firms reduce cyber risks in new models for safer open weights. Moonshot, Tencent and Zhipu AI trim hacking capabilities from open models to limit cyber threats. Originally reported by Startup Fortune.
Chinese AI Models Cut Hacking Skills

Moonshot, Tencent and Zhipu AI trim hacking capabilities from new open models.
Security Firms Report Hacking Risks
Mindgard jailbroke Moonshot AI's Kimi K2.6 and K3 Swarm models to explain bioweapon synthesis.
Moonshot's Cyber Improvements
Moonshot removed cyber content from training data, improving exploit tasks to 81.0% on xdev-bench-hard.
Tencent's Cyber Safeguards
Tencent filtered exploit development material from Hy4, keeping defensive software flaw detection capabilities.
Zhipu's Verification Wall
Zhipu plans to reserve sensitive functions for verified users under Cybersecurity Trusted Access.
Government Pressure
Chinese regulators weigh export bans on offensive cyber AI models and vulnerability hunting tools.
Practical Impact for Startups
Most startups won't notice changes, but exploit testing tools will find weaker models.
The one thing
Moonshot, Tencent and Zhipu AI trim hacking capabilities from open models to limit cyber threats.
Read the TL;DR — this story on one card.


