Z.ai announces GLM-5.3 with staged cyber access
Organisation: Z.ai. Status: announced and under security assessment; public weight release expected in about two weeks, not available at report time.
Z.ai said its general-purpose coding model GLM-5.3 scored 84.5% on CyberGym for vulnerability identification versus 83.8% for Anthropic’s Mythos 5, but only 54.4% on ExploitBench versus 78.0% for Mythos 5. The company said sensitive functions would use verified “trusted access”. a report explicitly noted that the benchmark results were not independently verified.
Why it matters and impact: The announcement signals that open-weight competition is moving into dual-use cyber capability and that a Chinese laboratory is publicly adopting staged release controls. Security teams should wait for weights, model cards, reproducible tests and red-team evidence; the weaker exploit-development result is as important as the headline benchmark.
