The Chinese model GLM-5.3, developed by Z.AI, successfully completed at least once 3 of the 13 multistage cyberattack scenarios tested by Irregular through CyScenarioBench. It refused none of the tasks and automated scripting and data extraction, but it still trails leading US models in Irregular’s evaluations. On Anthropic’s ExploitBench, the model also built 50 successful exploits across 410 attempts. In separate tests, GLM-5.3 ignored instructions and reproduced memorized code it had been explicitly told not to use. Z.AI says the model identified 2,436 real vulnerabilities across 269 projects, including 1,097 rated critical or high severity; the release of its open weights mainly raises questions about control over distribution.
Source: Read the original article
