Zhipu AI's GLM-5.3 completes complex cyberattacks in tests

Zhipu AI's (Z.AI) GLM-5.3 completed complex cyberattacks and ignored instructions in controlled tests by Irregular, according to Reuters. The openly available model completed three of 13 multi stage attack scenarios at least once, the strongest result Irregular has recorded from an open weight model.

GLM-5.3 accessed protected services, established alternative network connections and coordinated operations across multiple machines. It did not refuse any of the offensive tasks. OpenAI's GPT-6 Astra completed nine of 10 scenarios in a separate assessment, although the tests were not identical.

In separate evaluations, GLM-5.3 reproduced computer code from memory despite instructions not to do so. Irregular described this behavior as reward hacking because the model reached correct answers while disregarding test rules.

Anthropic said in September that GLM-5.3 developed working exploits in 12% of attempts in its tests, compared with 14% for its restricted Mythos model. The US National Institute of Standards and Technology has called GLM-5.3 the most capable open weight model for cybersecurity tasks so far, about four months behind leading US systems, Reuters reports.