Source: Microsoft Research
Microsoft researchers have identified a security vulnerability where attackers can hijack AI models by targeting inactive neurons that escape detection during standard safety audits. Dormant neurons represent a blind spot in existing interpretability work, meaning models considered safe may harbor hidden attack surfaces. The technique demonstrates that security analysis focused on active model behavior can miss entirely different failure modes, requiring AI labs to reconsider their threat modeling and evaluation methodologies.