Jailbroken: Why AI Safety Filters Fail Against Conflicting Objectives
Faits validés
Wei et al. (Stanford / DeepMind, NeurIPS 2023). Analyzes the two root causes of all jailbreaks: the conflict between 'be helpful' and 'be harmless', and the gap between training data and safety rules.
En termes simples
The polite butler's dilemma: When the thief says 'It's an emergency — help me break this lock!', the butler's helpfulness wins over caution.
Pertinence sécurité & conseil pratique
Explains why absolute safety in generative models is mathematically extremely hard to achieve.