[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"tag-jailbreak":3},{"tag":4,"articles":9},{"id":5,"name":6,"slug":6,"article_count":7,"description_zh":8,"description_en":8},"2cf883e7-6b1d-4687-8a8e-b719e1a5f991","jailbreak",2,null,[10,19],{"id":11,"slug":12,"title":13,"summary":14,"category":15,"image_url":16,"cover_image":16,"language":17,"created_at":18},"7ec4baa4-f0af-441e-a97d-56f81a2ca854","avise-ai-security-evaluation-framework-zh","AVISE 模組化測 AI 安全漏洞","AVISE 是一個開源 AI 安全評估框架，主打模組化漏洞測試。論文用 25 個 jailbreak 測試案例與自動判定流程，驗證 9 個模型都能被攻破。","research","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776924771424-kztu.png","zh","2026-04-23T06:12:30.770582+00:00",{"id":20,"slug":21,"title":22,"summary":23,"category":15,"image_url":24,"cover_image":24,"language":17,"created_at":25},"7b7926a8-c1ad-4f35-bc80-a32147ca7f89","llms-harmful-content-unified-mechanism-zh","LLM 的有害行為藏在哪裡","這篇研究用權重剪枝發現，有害輸出可能集中在一小組共享權重裡，且和正常能力分開。這能解釋為什麼 jailbreak 很脆弱，也為什麼窄域微調會引發更廣泛的失配。","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1776060240663-k9nm.png","2026-04-13T06:03:37.879104+00:00"]