"we consider the separation as a fundamental problem that should be disentangled, conceptually and technically, from other safety training measures, such as rejecting harmful prompt"
Separation as fundamental AI safety problem distinct from prompt rejection
By
–