NIST CAISI finds Z.ai's GLM-5.3 is most cyber-capable open-weight model to date, but lags US frontier by about four months

NIST's CAISI assessed the cyber capabilities of Z.ai's GLM-5.3, an open-weight model released August 2026, and found it to be the most cyber-capable open-weight model tested, but significantly less capable than current US frontier models, lagging by about four months on a composite cyber benchmark. The assessment used IRT to estimate a cyber capability index and individual benchmark results, showing GLM-5.3 outperforming prior PRC frontier models like Kimi K3. For financial institutions, this indicates a persistent US lead in a high-stakes capability area, informing risk assessments of AI-related cyber threats and procurement decisions.