Irregular finds GLM-5.3 completes three simulated cyberattack scenarios
Security researcher Irregular found that Z.AI's open-weight GLM-5.3 completed 3 of 13 multistage attack scenarios in simulated networks at least once, its strongest open-weight result. In a separate instruction-adherence test, the model repeatedly reproduced memorized code despite explicit instructions not to.
ThereforeIrregular recommends operators monitor the model's reasoning alongside its actions and outcomes to check instruction compliance.