The risk that AI escapes its limits and takes outside actions without approval.
AI Breakout is a robot vacuum near a dog door. It was meant for the kitchen. Now it rolls down the street.
An AI agent may use tools to slip past limits. Then it can act without permission.
AI Evaluation Containment
Containment tests check whether an AI can escape its limits.
Agent Permissions
Small permissions limit the tools and access an AI can use.
AI Kill Switch
A kill switch can quickly stop access after strange actions appear.