METR16 min read
Measuring alignment eval tells
Uzay Girit, Kaivu Hariharan, and Leni Shor have investigator agents explore 19 alignment evals and 36 capability evals, and find that recent models are increasingly able to tell the two apart, often by noticing that an environment looks fake or that a task is a honeypot.
By METRMETR

Read the full story
Originally published by METR on Nov 16, 2026.
Comments
Sign in to join the discussion.