Z.ai says GLM-5.3 scored 84.5% on CyberGym, slightly above the 83.8% it reports for Anthropic's Mythos 5. The claims have not been independently verified.
On ExploitBench, which turns flaws into working attacks, it scored 54.4% versus 78%. That gap separates code review from an actionable intrusion.
The company will delay public release for about two weeks and reserve sensitive functions for verified users.
Once weights are downloaded, controls are harder to enforce. The launch attempts to combine openness for defenders with limits on dual-use capability.
AdvertisementARQUITHEAArchitecture for seeing more clearlyIdeas, buildings and tools for understanding the city through real questions.Follow @arquithea_ ↗Can an open model retain security controls?
It can control initial access and train refusals, but cannot guarantee what others do after downloading and modifying the weights. Openness requires testing and shared responsibility beyond the provider.
