[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-kernel-level-safety-for-ai-tool-calls":10,"sections":34},{"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":29,"feedback":33,"feedback_at":22,"cost_usd":33,"total_tokens":33},4217,"kernel-level-safety-for-ai-tool-calls","Kernel-Level Safety for AI Tool Calls","A researcher built OS-level governance for MCP tool calls, arguing that app-layer guardrails are too easy to bypass.","AI agents can now make system calls that delete files, hit APIs, and touch shared state — and most safety checks live in software any script can sidestep.\n\nA paper published on arXiv proposes Governed MCP, a kernel-resident gateway that intercepts every tool call an AI agent makes through the Model Context Protocol. The system runs a six-layer pipeline — schema validation, trust tiers, rate limiting, an adversarial pre-filter, a semantic gate called ProbeLogits, and a constitutional policy check — all backed by a Blake3-hashed audit chain. It is implemented inside Anima OS, a bare-metal x86-64 kernel written in roughly 286,000 lines of Rust. The ProbeLogits gate classifies each call by running a probe prompt through a small model and reading a single logit, costing 332 to 556 milliseconds across Qwen2.5-7B, Llama-3-8B, and Mistral-7B — 2.4 to 3.4 times faster than a Llama Guard 3 pass on the same hardware.\n\nThe key finding is not the speed but the ablation: strip out ProbeLogits and F1 drops from 0.789 to 0.357, meaning hand-written firewall rules alone catch barely half of what they should. That is the argument for moving governance into the kernel — where a 10-line userspace script cannot route around it — rather than stacking more application-layer filters on top of each other.\n\nThe paper benchmarks at 98 to 99 percent on HarmBench and 98.5 to 100 percent unsafe recall on XSTest, numbers that look strong until you note the author flags a set of ring-3 syscall paths that remain ungated. Shipping a security primitive with known gaps is not unusual in research, but it is the kind of detail that tends to get buried in the headline numbers.","[\"ai\",\"security\",\"dev-tools\",\"open-source\"]","2026-07-07T04:00:00.000Z","2026-07-07T20:11:13.304Z","2026-07-07T20:11:16.243Z","published",null,[],"ai",[24,26,27,28],"security","dev-tools","open-source",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.16870",0,{"sections":35},[36,40,44,49,54,59,64,69,74,77,82,86,91,96],{"name":37,"slug":24,"count":38,"latest_published_at":39},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":41,"slug":26,"count":42,"latest_published_at":43},"Security",294,"2026-07-15T19:59:48.000Z",{"name":45,"slug":46,"count":47,"latest_published_at":48},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":58},"Hardware","hardware",122,"2026-07-14T19:46:26.000Z",{"name":60,"slug":61,"count":62,"latest_published_at":63},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":65,"slug":66,"count":67,"latest_published_at":68},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":70,"slug":71,"count":72,"latest_published_at":73},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":75,"slug":27,"count":76,"latest_published_at":18},"Dev Tools",59,{"name":78,"slug":79,"count":80,"latest_published_at":81},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":83,"slug":84,"count":80,"latest_published_at":85},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]