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

In three linesStudy on controlled removal of safety alignment in language models to evaluate cybersecurity capabilities. Compares authorized-context prompting, refusal-direction projection, and LoRA-based de-alignment. On 60 tasks (Security-AR), task-only LoRA reaches 0.87 security score with 0.83 general capability, but increases out-of-scope unsafe compliance.

Summary generated by Claude — human-verified