Chinese AI model carries out complex cyberattacks in controlled tests, researchers say

The GLM-5.3 model successfully completed 3 of 13 multi-stage cyberattack scenarios at least once, the strongest performance Irregular has recorded from an openly available AI model


