AI agents are increasingly used to complete complex tasks by interacting with external environments. These interactions can lead to unexpected behavior and expose agents to security threats, notably indirect prompt injection, where malicious instructions are hidden in external data.
To address this, the paper introduces RATION, a framework for quantitative risk-based access control. Instead of relying on static permissions, RATION aims to assess the risk of each action or interaction in a quantitative way, allowing agents to make safer decisions.
The approach is designed to mitigate the risks associated with external interactions by incorporating risk assessment into the access control mechanism. This could help prevent agents from being manipulated into performing harmful actions, though the paper is positioned as a step toward this goal rather than a complete solution.