Z.ai holds back GLM-5.3 weights after cyber benchmark results

In one sentence For the first time Z.ai delays an open-weights release: GLM-5.3, launched August 14, scored 84.5% on the CyberGym vulnerability-discovery benchmark, and downloadable weights stay restricted to vetted security partners until late August.

Needs review Reputable source

ShareLinkedIn X

Chinese lab Z.ai launched its new flagship model GLM-5.3 on August 14, but made an unprecedented call: the model weights, the file anyone can download to run it themselves, are not shipping right away. They will stay restricted for about two weeks to a small group of security partners, while the model remains usable through the API.

The reason lies in internal testing: GLM-5.3 turned out to be remarkably good at finding software vulnerabilities, to the point of beating the flagship models from Anthropic and OpenAI on one specific benchmark. A model this capable helps defenders, but once weights are public anyone can run it without filters, including to hunt for exploitable flaws.

It is the first time Z.ai, historically among the fastest labs to publish open weights, chooses to slow down for safety reasons. The move echoes OpenAI, which in early August slowed the release of its Astra model citing offensive cyber capabilities. The topic is no longer theoretical: models are becoming concrete instruments in the attack-defense game.