Overview
- Z.ai announced GLM-5.3 on Friday and published company-reported results showing gains in coding and cyber-defence benchmarks compared with its prior release.
- The company reported an 84.5% score on CyberGym for finding and confirming software flaws and a lower 54.4% on ExploitBench for turning those flaws into working attacks, numbers that have not been independently verified.
- Z.ai said it will release the model’s weights, the downloadable parameters developers use to run and modify the system, in about two weeks after completing security assessments.
- For sensitive cybersecurity tasks Z.ai introduced a ‘trusted access’ programme that limits features to verified users and added layered safeguards meant to screen risky requests monitor model outputs and train the model to refuse malicious prompts.
- Z.ai also published a ledger listing 2,436 security findings across open-source projects and framed GLM-5.3 as a challenge to restricted-access tools which raises questions about dual-use risks and how defenders and policy makers balance openness with safety.