Ablating Safety: Mechanisms for Removing Alignment in Language Models for Security Applications | ArxivCSExplorer