MIRI researchers identify security gaps in proposed international AI chip monitoring agreement designed to prevent unsafe superhuman AI development.

MIRI's Technical Governance Team proposed an international agreement requiring AI chip clusters with 16+ H100 GPU equivalents to register with a US-China-led coalition for safety oversight.

The proposal aims to halt risky superhuman AI development until safety measures are adequate, with mechanisms to monitor cluster operations.

A critical vulnerability exists in the agreement's definition of covered chip clusters, leaving potential loopholes for illicit distributed training operations to evade detection.

The authors acknowledged various contingencies and attempted to close gaps, but the threat model for unregistered or distributed computing arrangements was insufficiently addressed.