Overview
- Z.ai announced GLM-5.3 on Friday, Aug. 14 and made the model available through subscription endpoints while saying it will publish downloadable weights in about two weeks after completing safety hardening.
- Z.ai reports GLM-5.3 scored 84.5% on CyberGym, a benchmark for identifying real software flaws, and 54.4% on ExploitBench, which measures turning flaws into working attacks.
- The company says the gains came from expanded post-training on long-horizon coding and security tasks rather than changes to the 743‑billion‑parameter base model.
- Independent researchers and Z.ai flagged real findings from the model, including a serious vulnerability in the Cursor code editor, and Z.ai published a ledger listing 2,436 reported security findings across 269 open‑source projects.
- Z.ai plans layered safeguards and a 'trusted access' path for its most sensitive functions but outside experts warn those protections will be harder to enforce once weights are downloadable and that the release raises dual‑use, market and geopolitical risks.