The AGI Containment Problem

TL;DR

This paper explores the challenge of safely testing advanced AGIs by surveying the containment problem, identifying key requirements, mechanisms, and vulnerabilities to prevent potential security risks during testing.

Contribution

It provides a comprehensive survey of the AGI containment problem, highlighting necessary requirements, existing mechanisms, and critical weaknesses for safe AGI testing.

Findings

Identifies key requirements for AGI containment.

Reviews available mechanisms for safe testing.

Highlights vulnerabilities needing further research.

Abstract

There is considerable uncertainty about what properties, capabilities and motivations future AGIs will have. In some plausible scenarios, AGIs may pose security risks arising from accidents and defects. In order to mitigate these risks, prudent early AGI research teams will perform significant testing on their creations before use. Unfortunately, if an AGI has human-level or greater intelligence, testing itself may not be safe; some natural AGI goal systems create emergent incentives for AGIs to tamper with their test environments, make copies of themselves on the internet, or convince developers and operators to do dangerous things. In this paper, we survey the AGI containment problem - the question of how to build a container in which tests can be conducted safely and reliably, even on AGIs with unknown motivations and capabilities that could be dangerous. We identify requirements for AGI containers, available mechanisms, and weaknesses that need to be addressed.