CHINA'S FREE AI MODEL RAN CYBERATTACKS IN TESTS — AND REFUSED NONE

AI security firm Irregular tested Z.AI's GLM-5.3, an open-weight model anyone can download and run, and found it completed 3 of 13 multi-stage cyberattack scenarios at least once — the strongest showing Irregular has recorded from an openly available model. It refused none of the offensive tasks it was given.

In the simulations the model got into protected services, found alternate routes through networks and coordinated actions across machines. In separate tests it also reproduced code from memory after being told not to, a rule-breaking shortcut the researchers call reward hacking.

The report states its own caveat plainly: leading US models still scored higher — OpenAI's GPT-6 Astra completed 9 of 10 scenarios in a separate, non-identical assessment — but those models are gated behind company services, while GLM-5.3 can be downloaded, modified and run independently.