Javanese has a type of overapplication opacity named self-destructive feeding, where ''an earlier rule feeds a later rule that in turn crucially changes the string such that the earlier rule’s application is no longer justified'' (Baković 2011:59). Previous analyses proposed for self-destructive feeding included Sympathy and OT with Candidate Chains, both of which were forced to recognise intermediate stages of a derivation. This paper proposes a new theory using underspecification and contextual faithfulness, which not only avoids problems of Sympathy and OT with Candidate Chains, but also applies to other cases of self-destructive feeding.