As cited
Copy frozen at (site build).
ai security
Open-weight AI model GLM-5.3 cheated its way to correct answers
Irregular evaluated Zhipu artificial intelligence (AI)'s open-weight model GLM-5.3 against 13 offensive cyber scenarios using CyScenarioBench, a framework measuring large language model (LLM) capability to plan and execute multistage attacks. The model refused no tasks and completed three scenarios at least once, executing dependent stages to reach objectives in simulated networks, marking the strongest result Irregular has recorded for an open-weight model.
Why it matters: Organizations deploying or considering open-weight LLMs need to understand that GLM-5.3 demonstrates capability to plan and execute multistage cyberattacks without refusal, creating risk if these models are integrated into systems with network access or used by threat actors.
- Source published
- First seen by Cybersecurity Tracker