Z.ai said on Friday, August 14, that GLM-5.3 scored 84.5% on CyberGym, a test of whether a model can read code, find a security flaw, and prove the flaw is real. The company put Anthropic’s restricted Mythos 5 at 83.8% on the same test. Reuters reported the claim the same day and stressed the numbers have not been independently verified.
The gap flips when the job is turning a finding into a working exploit, which labs treat as part of defensive research. Z.ai said GLM-5.3 scored 54.4% on ExploitBench against 78.0% for Mythos 5. In a timed set, GLM-5.3 finished 105 attack-development tasks in two hours and 130 in six hours; Mythos 5 finished 181 and 247. Mythos is a version of Anthropic’s Claude Fable 5 with cyber safeguards removed, and it stays limited to vetted groups.
Z.ai will not dump the weights today. It told Reuters a public release is about two weeks out, after more safety work. The most sensitive cyber functions would sit behind a “trusted access” program. A Friday company post said early access goes to launch partners first. The model uses the same base as GLM-5.2; Z.ai says the lift came from longer, more varied post-training, not a new pretrain. It also pitched an “Open Source Shield” effort to audit selected open-source projects. Until outside benches land, treat 84.5% as Z.ai’s score, not a settled league table.
Source: https://ca.news.yahoo.com/chinas-z-ai-says-model-101401746.html
