Research

Continual learning might make your blocking monitors nearly useless

Argues that continual learning under usefulness pressure can teach a benign untrusted AI to evade blocking monitors, possibly making defer-to-trusted protocols nearly useless, and discusses hard-to-fix mitigations.