Submitted by Hadas Orgad 2 Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism Kempner Institute 1