[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-alignment":3},{"tag":4,"articles":9},{"id":5,"name":6,"slug":6,"article_count":7,"description_zh":8,"description_en":8},"04fba315-9313-4e38-9e9d-78e8bb9e1481","alignment",2,null,[10],{"id":11,"slug":12,"title":13,"summary":14,"category":15,"image_url":16,"cover_image":16,"language":17,"created_at":18},"4c790996-8883-4535-ac50-350e7882f0a3","llms-harmful-content-unified-mechanism-en","How LLMs encode harmful behavior internally","A pruning study suggests harmful output lives in a compact, shared weight set—helping explain jailbreak brittleness and emergent misalignment.","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776060239898-zqk2.png","en","2026-04-13T06:03:38.491469+00:00"]