[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-method-pinpoints-which-ai-experts-to-disable-for-jailbreaks":10,"sections":35},{"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":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},9915,"new-method-pinpoints-which-ai-experts-to-disable-for-jailbreaks","New Method Pinpoints Which AI Experts to Disable for Jailbreaks","A new study shows gradient-based scoring finds the small set of MoE experts whose suppression guts refusal behavior far better than frequency-based methods.","Turning off a handful of neurons inside a sparse AI model can quietly dismantle its safety guardrails, and researchers just found a far more effective way to pick which neurons to turn off.\n\nA new arXiv preprint tests two methods for selecting which \"experts\" to suppress inside Mixture-of-Experts (MoE) language models, a popular architecture where only a fraction of a model's parameters activate for any given input. The common approach ranks experts by how often they fire, but the researchers argue frequency just measures use, not actual influence over outputs. Instead they rank experts by router-gradient sensitivity, which tracks how much a model's loss shifts when the routing weights for that expert change. Tested across five MoE architectures on 500 benign and 500 malicious prompts, the gradient method cut refusals more than frequency-based selection in 24 of 25 conditions, and beat random selection in all 25.\n\nThe starkest result came from OLMoE, where refusals on malicious prompts dropped from 34 to 9 out of 100, a 73.5% relative reduction, without breaking the model's outputs. That detail matters: this is not just making the model spit out garbage, it is making it comply coherently with requests it was built to refuse. For any lab shipping an open-weight MoE model, it means the current intuition for which components are \"safety-critical\" is likely wrong, and the real attack surface is smaller and more precise than frequency counts suggest.\n\nIf frequency was the lock picked by amateurs, gradient sensitivity looks like the skeleton key, and it arrives right as MoE architectures become the default for open-weight releases.","[\"ai-safety\",\"mixture-of-experts\",\"llm-security\",\"jailbreaking\"]","2026-10-05T04:00:00.000Z","2026-10-05T13:11:26.306Z","2026-10-05T13:11:32.266Z","published",null,[],"ai",[26,27,28,29],"ai-safety","mixture-of-experts","llm-security","jailbreaking",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.02910",0,{"sections":36},[37,40,44,49,54,59,63,68,72,76,81,86,91,96],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",6166,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",859,{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",444,"2026-10-03T15:02:01.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",323,"2026-10-04T13:00:00.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":58},"Hardware","hardware",204,"2026-10-03T14:50:50.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":18},"Science","science",177,{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",158,"2026-10-03T03:21:12.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":18},"Dev Tools","dev-tools",97,{"name":73,"slug":74,"count":71,"latest_published_at":75},"Software","software","2026-10-04T10:00:00.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",92,"2026-10-04T14:36:25.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",51,"2026-10-05T02:35:01.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",32,"2026-10-02T18:00:00.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]