Ai safety framework enforces action limits through monitoring roles
Contract monitoring: governing AI via separation of powers
Artificial Intelligence
Summary
AI systems can sometimes do things they shouldn’t, which can cause safety problems. This paper presents a way to keep AI workers in check by giving them contracts that limit what they can do. Special monitors and judges watch over the workers and make sure they follow these contracts using different computing powers. This lets us measure how safe the system is in a clear way, and the approach works in many situations, including software security and keeping AI contained.
What this means in practice
- •For software security teams: Implement layered monitoring to ensure code actions comply with safety contracts, improving trusted execution monitoring.
- •For robotics control engineers: Use the framework to limit autonomous robot behaviors by enforcing action boundaries monitored by specialized agents.
Authors
Enric Boix-Adsera
Abstract
We propose an AI safety framework that binds worker agents to contracts specifying their permitted actions. We show how these contracts can be enforced and specified by assigning distinct responsibilities to monitor agents and judges, and asymmetric computational resources to monitors and workers. Our framework allows us to empirically measure statistical safety guarantees. The framework applies to a wide range of settings, including code security and escape-the-box scenarios.