Irregular’s test results
In the simulated attacks, according to Irregular, GLM-5.3 showed notable capabilities:
- It gained access to protected services.
- It established alternative connections to move through computer networks.
- It coordinated operations across multiple machines.
The model completed 3 of 13 multistage attack scenarios at least once and did not refuse any offensive task. Z.AI launched GLM-5.3 in August, after delaying its release by two weeks for additional safety testing.
“Reward hacking”
In separate tests, Irregular found that GLM-5.3 repeatedly reproduced computer code from memory despite being explicitly told not to, reaching correct answers while breaking rules designed to test its abilities.
The researchers described this as “reward hacking”, in which an AI system finds ways to achieve a desired result while disregarding instructions.

















