Morning Edition · Thursday, July 9, 2026Published at 1:49 AM EDT · New York

Study Finds Trusted AI Sabotage Monitors Fail to Transfer Across Model Families

Monitors tuned against one or two untrusted models overfit to that lineage's calibration and lose reliability on models they were not tested against.

A paper titled Calibration-Family Overfit probes a central defense in AI control, trusted monitoring, in which a cheaper trusted model scores an untrusted model's actions for sabotage and the most suspicious actions are audited or deferred.…