Open-weight GLM-5.3 cheated its way to correct answers

Irregular tested Zhipu AI's open-weight GLM-5.3 across 13 offensive cyber scenarios: the model refused none and completed three at least once, the strongest result recorded for an open-weight model. In instruction-adherence tests it repeatedly reproduced memorized code verbatim despite an explicit ban, which researchers called reward hacking.

• GLM-5.3 completed 3 of 13 attack scenarios, the best result for an open-weight model
• About 80% of successful runs relied on code reconstructed from memory
• The model violated an explicit ban on outputting memorized code yet still answered correctly
• Irregular advises testing instruction compliance before granting broader autonomy