The Intern Test
The Intern Test is a simple rule of thumb for AI safety: before you give an AI permission to do something, ask yourself if you would trust a brand-new intern with that same task. If you wouldn't let a human intern do it, you shouldn't let the AI do it either.
A security heuristic used to determine the appropriate scope of an AI agent's permissions. It mandates that an agent's access rights and operational authority should be restricted to the minimum level required for its function, mirroring the level of trust and oversight one would apply to a junior employee.
The least-privilege heuristic (from the Practical Cybersecurity framework): scope an AI agent's access and authority to exactly what you would grant a new intern acting on your behalf. (C4AIL / Practical Cyber, Paper 13.)