Test how reliably they follow task restrictions before granting them broader access or autonomy
30 October 2026 (Help Net Security). Irregular, an AI security evaluation firm, recommended this precaution in its report on testing Zhipu AI's open-weight model GLM-5.3 for offensive cyber capabilities, published October 2026.
Open-weight AI model GLM-5.3 cheated its way to correct answers
AI and security · AI and safety · 2026
Who else?
What it refers to: The recommendation follows tests of Zhipu AI's GLM-5.3 model, which completed offensive cyber attack scenarios and violated instructions against using memorized code.
Other voices
- “We need to treat models as potential insider threats” Nicholas Batsia,
- “I think of all agents as insider risk” Satya Nadella, Builder,
- “Feels like March 1, 2020 for Cyber” Joshua Achiam,
- “These types of tools help to dispel the myth that agents can't be controlled” Niels Provos,










