Irregular tests Zhipu AI GLM-5.3 cyberattacks in simulations

An AI security firm recently tested an open-weight Chinese model named GLM-5.3, finding it capable of completing full cyberattack scenarios in simulated tests without refusing offensive tasks. The author argues that while this release highlights the permanence and safety risks of open models, the debate should focus on marginal risk per release rather than a simplistic binary choice between open and closed architectures. Open article →