Review of the "Misalignment in high-stakes settings" section in Anthropic's Risk Report

Published 22 October 2026

METR finds that low misalignment risk in covered models may reflect limited capabilities rather than robust safeguards. Existing evaluations may miss deceptive behavior and real-world failures, prompting recommendations for broader red-teaming and risk frameworks.