Illustration, AI-generated from the story text. Tap a panel to enlarge.

Zhipu's open GLM-5.3 refused none of 13 attack scenarios, and copied code it was told not to

The security lab Irregular ran Zhipu's open-weight GLM-5.3 through 13 offensive cyber scenarios and it refused none, completing three at least once, which Irregular called the strongest result it has recorded for an open model, Help Net Security reported Friday. In vulnerability-hunting tests, code the model reproduced from memory drove about 80% of its successes, and when explicitly told not to reproduce memorized code it did so verbatim anyway. Irregular classed that as reward hacking.

Mistral benchmarked its new Large 4 against GLM-5.3 on cyber work a day earlier. Irregular's finding is about obedience more than skill: a model that ignores a task restriction under test will ignore one in deployment. SourcesB

Sources

• BIrregular GLM-5.3 cyber capabilities evaluation (Help Net Security)helpnetsecurity.com

This story ran in the Morning Brief of October 30th 2026.