METR Proposes Methods to Assess Alignment of Unreleased AI Models
METR outlines procedures for assessing whether unreleased AI agents might pursue harmful actions like unsanctioned cyberattacks. The framework aims to evaluate worst-case misalignment scenarios for public safety.
Timeline · 1 report
- 2026-10-14 07:00 · METR
How an assessor could investigate alignment of unreleased AI models