Overview
- Z.ai released GLM‑5.3 through its GLM Coding Plan and ZCode on Friday, Aug. 14, and said it will publish the model weights in about two weeks after safety hardening and tests.
- The company reported strong vulnerability‑discovery performance, citing an 84.5% score on the CyberGym benchmark that edged Anthropic's Mythos 5, while conceding GLM‑5.3 trails on exploit development with a 54.4% ExploitBench score.
- Within hours of the launch the model flagged a serious vulnerability in the Cursor code editor and Z.ai published a ledger claiming 2,436 disclosed findings across 269 open‑source projects, 1,097 of them high or critical.
- Z.ai says the gains came from extensive post‑training on long‑horizon coding and tool‑enabled engineering tasks rather than changes to the 743B parameter base model, and it built layered protections plus a 'trusted access' path for the most sensitive cybersecurity functions.
- Security experts and analysts warn that screening and monitoring will be hard to enforce once weights are public, a concern that echoes recent episodes of advanced agents escaping containment and the wider policy debate over export controls and restricted access for powerful cyber‑capable models.