[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-gating-method-fixes-blind-spot-in-ai-tool-agents":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},6126,"new-gating-method-fixes-blind-spot-in-ai-tool-agents","New Gating Method Fixes Blind Spot in AI Tool Agents","A new admission method for AI tool agents blocks a flaw that let scalar safety checks approve locally better but ultimately unsafe actions.","A new paper identifies a structural flaw in how AI agents get checked before they act, and offers a fix that closes it completely in testing.\n\nThe researchers built a system called SiLR that changes how a gate, the runtime check that decides whether an AI tool agent's next action gets approved, works. Most gates today score a proposed action with a single number and approve it if the score improves. The paper shows that approach can get an agent stuck accepting locally better moves that never actually fix the underlying problem, a failure the authors call the scalar projection trap. SiLR instead simulates each proposed action first and checks it against multiple violation states at once, rather than collapsing everything into one score. In tests on power grid simulations, SiLR recovered all 21 multi-action episodes it was tested on, compared to zero for the simplest gate and nine for the best single-score alternative.\n\nThat gap matters because single-score safety checks are a common shortcut in agent design, and this paper argues they are not just poorly tuned but structurally unable to represent the problem, no matter how the threshold is adjusted. The same full check also blocked a deliberate attack that fooled both the single-score gate and a simpler multi-branch check, and when reused as a training signal it produced a policy that improved on the untrained baseline, unlike its simpler counterpart.\n\nIt is a reminder that in safety-critical agent systems, one number is rarely the whole story, and corner-cutting on gating logic tends to get caught eventually.","[\"ai-agents\",\"ai-safety\",\"llm\",\"reinforcement-learning\"]","2026-09-07T04:00:00.000Z","2026-09-07T05:17:49.081Z","2026-09-07T05:18:01.125Z","published",null,[],"ai",[26,27,28,29],"ai-agents","ai-safety","llm","reinforcement-learning",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.04629",0,{"sections":36},[37,40,44,49,54,59,64,69,74,79,84,89,94,99],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",3399,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",568,{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",310,"2026-09-06T15:34:34.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":58},"Hardware","hardware",152,"2026-09-03T09:26:48.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":63},"Consumer Tech","consumer-tech",97,"2026-09-04T15:29:18.000Z",{"name":65,"slug":66,"count":67,"latest_published_at":68},"Science","science",96,"2026-09-03T22:30:00.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Software","software",73,"2026-08-18T07:51:50.000Z",{"name":75,"slug":76,"count":77,"latest_published_at":78},"Dev Tools","dev-tools",69,"2026-08-18T04:00:00.000Z",{"name":80,"slug":81,"count":82,"latest_published_at":83},"Startups","startups",54,"2026-09-04T23:36:14.000Z",{"name":85,"slug":86,"count":87,"latest_published_at":88},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":90,"slug":91,"count":92,"latest_published_at":93},"General","general",39,"2026-09-05T16:13:57.000Z",{"name":95,"slug":96,"count":97,"latest_published_at":98},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":100,"slug":101,"count":102,"latest_published_at":103},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]