[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-system-vets-every-ai-agent-action-before-it-runs":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},8521,"new-system-vets-every-ai-agent-action-before-it-runs","New System Vets Every AI Agent Action Before It Runs","A deterministic policy engine for enterprise AI agents blocked every unsafe action in testing, but stayed more cautious than human reviewers preferred.","A new governance system checks every action an AI agent wants to take, and in testing it never let a dangerous one through.\n\nCalled VeriWeave Govern, the system splits AI decision-making from AI permission-granting: an agent proposes an action, and a separate deterministic layer checks it against versioned policies and typed evidence before anything executes. Actions get sorted into deny, review, or allow, in that strict order, and anything consequential gets routed to a human. Researchers tested the design on GovernBench, a benchmark of 60,000 labeled cases across five enterprise domains, including adversarial evidence and shifting policies over time, run across 30 seeds. The system hit 0.9888 accuracy, 0.9836 macro-F1, and recorded zero false allows and zero successful attacks across the entire test set, plus a clean run on 12 end-to-end scenarios and a 40,040-request load test.\n\nThe more telling result comes from a smaller test: 150 cases grounded in actual EU and Austrian regulation, judged by two human annotators who agreed on every call. There, the deterministic engine played it safe more often than the humans wanted, while a general-purpose model, Gemma 4 31B, tracked the human verdicts more closely. That is the real finding here: airtight safety and good judgment are not automatically the same thing, and enterprises rolling out agents that touch real infrastructure will have to pick a point on that tradeoff.\n\nA system that never once approved something bad, in a test it designed itself, is reassuring right up until it also refuses things a human would have waved through.","[\"ai-agents\",\"ai-governance\",\"enterprise-ai\",\"ai-safety\"]","2026-09-30T04:00:00.000Z","2026-09-30T08:44:16.080Z","2026-09-30T08:44:22.142Z","published",null,[],"ai",[26,27,28,29],"ai-agents","ai-governance","enterprise-ai","ai-safety",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.37457",0,{"sections":36},[37,40,44,48,53,58,63,68,73,78,83,88,93,98],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",5028,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",780,{"name":45,"slug":46,"count":47,"latest_published_at":18},"Policy","policy",417,{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",284,"2026-09-29T21:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":57},"Hardware","hardware",194,"2026-09-29T13:16:04.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Science","science",154,"2026-09-28T13:19:18.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":67},"Consumer Tech","consumer-tech",142,"2026-09-29T18:38:03.000Z",{"name":69,"slug":70,"count":71,"latest_published_at":72},"Software","software",91,"2026-09-25T20:55:00.000Z",{"name":74,"slug":75,"count":76,"latest_published_at":77},"Dev Tools","dev-tools",89,"2026-09-29T17:15:00.000Z",{"name":79,"slug":80,"count":81,"latest_published_at":82},"Startups","startups",83,"2026-09-29T21:51:36.000Z",{"name":84,"slug":85,"count":86,"latest_published_at":87},"General","general",49,"2026-09-28T16:44:57.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]