GLM-5.3 announced with delayed weights and staged access for cyber capabilities

Z.ai announced GLM-5.3 on August 14, 2026. At the report cutoff there were no public weights, and a security review was under way.

The company claimed 84.5 per cent on CyberGym against 83.8 per cent for Anthropic's Mythos 5, but on ExploitBench GLM-5.3 scored 54.4 per cent against Mythos 5's 78.0 per cent. The results have not been independently verified. Sensitive functions are planned to be provided to verified users through 'trusted access'.

Why it matters: open-weight competition is extending into dual-use cyber capability. Organisations should wait for the weights, model card and reproducible evaluation, and implement identity, tool and telemetry controls in the meantime.