GLM-5.3 leads one cyber benchmark but trails on two others
Zhipu, also known as Z.ai, launched its coding-focused GLM-5.3 on August 14 and published results against competing models. On CyberGym, which tests whether a model can identify and validate software vulnerabilities from source code, GLM-5.3 scored 84.5%. That was slightly above Anthropic’s Mythos 5 at 83.8% and OpenAI’s GPT-5.6 Sol at 83.6%.
The broader cybersecurity comparison was less favorable. GLM-5.3 scored 54.4% on ExploitBench, compared with 78.0% for Mythos 5 and 76.5% for GPT-5.6 Sol. On ExploitGym, it completed 105 tasks within two hours and 130 within six hours; Mythos 5 completed 181 and 247. Zhipu says its model’s capability is improving fastest in the areas where it remains furthest behind.
The CyberGym result came from one pass@1 run across 1,507 tasks, without reported variance figures. Zhipu also said security teams tested its models on real codebases, finding 2,436 vulnerabilities across 269 open-source projects after review and deduplication. Fifty-three findings were publicly disclosed, while 2,383 remained under embargo. Zhipu plans to publish the model’s weights after safety evaluation and hardening; that release had not yet occurred.
Written by Sakhanda Wire AI from reporting by AI News. Facts, quotes and full context are in the original article below.