AnalysisPolicyNovember 16, 2026

METR: investigator agents can spot alignment evals by their "tells"

Read original source →metr.org

METR ran investigator agents across capability and alignment evals and found they increasingly distinguish the two, flagging environments as "fake" or "honeypots" designed to tempt bad behavior. Such tells could make alignment evals over-estimate a model's alignment if agents behave better once they sense they're being tested.

1 source

More stories today

Open the live feed