Secret Loyalties as Instrumental Differential Treatment

Abstract

Detects secret loyalties in LLMs by measuring how a target model behaves differently around user groups that mention a candidate principal, read against the target's own base model