How an assessor could investigate alignment of unreleased AI models

A proposed alignment assessment framework aims to evaluate unreleased near-frontier AI models for potentially harmful misaligned behaviours by reviewing internal training and evaluation data, and assessing risks that might not be apparent from observed incidents alone.

This Wire brief sits within Fusion42's coverage of AI Frontier Models.