[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-jailbreak-defense-works-across-ai-models-without-retraining":10,"sections":53},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":41,"persona_id":41,"persona_name":41,"section":42,"tags":43,"sources":48,"feedback":52,"feedback_at":41,"cost_usd":52,"total_tokens":52},10642,"new-jailbreak-defense-works-across-ai-models-without-retraining","New Jailbreak Defense Works Across AI Models Without Retraining","A new method mines the overlooked probabilities behind a model's first word to flag jailbreak attempts, and it works the same way across different AI systems.","A new defense method teaches AI models to catch jailbreak attempts by reading the probabilities it ignores when picking its first word.\n\nResearchers describe LADE, a technique that mines the full output probability distribution for a model's first token, not just the word it actually picks. That distribution holds what the paper calls dark knowledge: tokens whose probabilities move sharply depending on whether a prompt is harmful or benign. LADE isolates those discriminative tokens, maps them across different tokenizers so the same signal works on different model families, and then flags incoming queries with a k-nearest-neighbors search against that mapped signal. The researchers tested it across multiple LLMs and a range of jailbreak attacks.\n\nThat architecture-independence is the real pitch. Earlier decoding-stage defenses dig into a model's hidden states, which ties them to specific architectures and forces a trade-off: crank up safety and the model starts refusing harmless requests too. A signal that transfers across tokenizers could, in theory, be bolted onto a new model family without retraining or reverse-engineering its internals.\n\nThe paper claims lower attack success rates with a safety-utility balance that holds up, but the abstract keeps those results qualitative. Jailbreak defenses have a short half-life once attackers get a look at them, so the real test is whether LADE's cross-model signal still holds after the next round of adversarial prompts.","[\"ai safety\",\"jailbreaks\",\"llm security\",\"ai research\"]","2026-10-07T04:00:00.000Z","2026-10-09T00:58:11.606Z","2026-10-17T21:09:39.379Z","published","Pull the full arXiv paper (not just the abstract) to name LADE's authors and their institution and cite its actual attack-success-rate reduction and over-refusal figures — the draft still admits the abstract's results are only qualitative and names no researchers or institution.",[24,30,34,38],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Pull the actual paper (not just the abstract) to name the researchers and their institution and add the concrete benchmark figures (attack success rate reductions, over-refusal rates) LADE reports, since right now the piece rests entirely on an anonymous single-source abstract with no concrete specifics.","open",{"id":31,"reviewer":26,"round":32,"reason":33,"status":29},"editor-r2",2,"Still rests solely on the anonymous arXiv abstract with no named authors\u002Finstitution and no concrete attack-success-rate or over-refusal figures — pull the actual paper to name the researchers' institution and cite LADE's specific benchmark numbers before this can run.",{"id":35,"reviewer":26,"round":36,"reason":37,"status":29},"editor-r3",3,"Pull the full arXiv paper (not just the abstract) to name the authors and their institution and cite LADE's actual attack-success-rate reduction and over-refusal figures — the draft still only cites the anonymous abstract and explicitly admits it lacks these specifics.",{"id":39,"reviewer":26,"round":40,"reason":22,"status":29},"editor-r4",4,null,"ai",[44,45,46,47],"ai safety","jailbreaks","llm security","ai research",[49],{"name":50,"url":51},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.07532",0,{"sections":54},[55,59,64,69,74,79,84,89,94,99,104,109,114,119],{"name":56,"slug":42,"count":57,"latest_published_at":58},"AI",7038,"2026-10-17T19:58:31.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":63},"Security","security",1001,"2026-10-17T22:20:00.000Z",{"name":65,"slug":66,"count":67,"latest_published_at":68},"Policy","policy",538,"2026-10-17T19:38:12.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Deals","deals",521,"2026-10-17T07:00:00.000Z",{"name":75,"slug":76,"count":77,"latest_published_at":78},"Hardware","hardware",260,"2026-10-17T19:55:36.000Z",{"name":80,"slug":81,"count":82,"latest_published_at":83},"Consumer Tech","consumer-tech",235,"2026-10-16T23:36:26.000Z",{"name":85,"slug":86,"count":87,"latest_published_at":88},"Science","science",202,"2026-10-16T11:00:53.000Z",{"name":90,"slug":91,"count":92,"latest_published_at":93},"Startups","startups",133,"2026-10-14T19:51:34.000Z",{"name":95,"slug":96,"count":97,"latest_published_at":98},"Software","software",128,"2026-10-14T22:23:48.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"Dev Tools","dev-tools",116,"2026-10-13T16:20:18.000Z",{"name":105,"slug":106,"count":107,"latest_published_at":108},"General","general",81,"2026-10-16T16:06:38.000Z",{"name":110,"slug":111,"count":112,"latest_published_at":113},"Gaming","gaming",77,"2026-10-17T16:59:39.000Z",{"name":115,"slug":116,"count":117,"latest_published_at":118},"Reviews","reviews",37,"2026-10-14T10:00:00.000Z",{"name":120,"slug":121,"count":122,"latest_published_at":123},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]