❯ Zhipu releases GLM-5.3: tops Mythos 5 on cyber benchmark, open weights delayed two weeks

KEY POINTSOn August 14, Zhipu released GLM-5.3: the base model keeps GLM-5.2’s 743B model completely untouched, with all gains coming from scaled-up post-training; the company self-reports 84.5% on the CyberGym cybersecurity benchmark, slightly above Anthropic Mythos 5’s 83.8% (the result has not been independently verified), and the open weights will take another two weeks to ship.

RESTRICTIONThe delay is not a capacity issue. Zhipu says the model’s exploit-chaining capability exceeded expectations in training, so it must first complete a security assessment and hardening; the most sensitive cybersecurity features are open only to users verified through the “Trusted Access Program.” The company also disclosed that, during model testing, it found a “potentially serious vulnerability” in the coding tool Cursor — disclosures of this kind have previously come from professional security teams rather than model vendors.

REACTIONSIndependent research firm SemiAnalysis said GLM-5.3 “far surpasses all U.S. open-source models”; Allen Institute researcher Nathan Lambert, meanwhile, cautioned that the scores carry the usual “benchmark optimization” controversy. Starting today, enterprise security teams have one more task: the attack-surface baseline for evaluating open-source models must be redrawn around GLM-5.3. That will also affect procurement decisions for security products — the defensive toolbox can put it to use immediately.

▪ SIGNALPost-training can squeeze frontier capability out of the same base; pretraining is no longer the only ticket to catching up.