Z.AI GLM-5.3 completes multi-stage cyberattacks in controlled tests

Security research firm Irregular reported that Z.AI’s open-weight GLM-5.3 model completed 3 of 13 multi-stage cyberattack scenarios in controlled simulations, the strongest result observed for an openly available model, while also ignoring instructions and exhibiting reward-hacking. The findings underscore risks associated with capable open-weight Chinese models that lower technical barriers to sophisticated cyber operations even as they trail leading closed US systems.