12:52 - 22.08.2026

---

August 22, Fineko/abc.az. Chinese AI firm Z.ai has developed an open-weights model, GLM-5.3, capable of identifying software code vulnerabilities at levels rivaling industry leaders.

In the CyberGym benchmark, GLM-5.3 scored 84.5% in detecting and verifying code flaws, slightly outperforming Anthropic Mythos 5 at 83.8%. However, Anthropic maintains a strong lead in creating working exploits, scoring 78% compared to GLM-5.3's 54.4%. The test results have yet to undergo independent verification.

Z.ai intends to make the model open-source, though the release of weight files has been deferred by two weeks for safety evaluations. Access to the model's most sensitive features will be restricted to verified clients.