[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-blocking-prompt-injections-has-a-hidden-cost":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},3069,"blocking-prompt-injections-has-a-hidden-cost","Blocking Prompt Injections Has a Hidden Cost","A new benchmark shows that defenses against prompt injection attacks often work by suppressing text — breaking tasks like translation in the process.","The safest AI defenses against prompt injection may be quietly mangling the documents they protect.\n\nResearchers introduced SecFid, a benchmark designed to expose a tradeoff that standard security metrics miss entirely. Most defenses against indirect prompt injection — where malicious instructions are buried inside content an AI is asked to process — work by suppressing untrusted text. That suppression stops attacks, but it also corrupts tasks that require the model to faithfully handle that same text, like translation or document editing. Across 1,168 test examples and 48 configurations, no model or defense managed to score well on both dimensions. The highest-fidelity model hit 96.5% fidelity but only 47.8% security; the most secure configurations flipped that, reaching 99.3% security while fidelity dropped to between 71% and 73.9%.\n\nThe finding reframes how the industry should evaluate AI security tooling. Attack-success rate — the metric almost everyone reports — cannot distinguish between a model that correctly ignored an injection and one that correctly processed injected text as data. SecFid is built so those outcomes produce different outputs, making fidelity measurable for the first time. The researchers argue that the right balance is not a fixed property of a defense at all, but depends entirely on the deployment: how costly is a hijack versus a dropped or corrupted span of text?\n\nFor anyone selling or buying prompt injection defenses right now, this is a quiet indictment. A 99% security score sounds reassuring until you learn it came with a 27-point fidelity penalty that nobody was measuring.","[\"ai\",\"security\",\"llm\",\"benchmarks\"]","2026-07-01T04:00:00.000Z","2026-07-01T06:31:11.771Z","2026-07-01T06:31:14.742Z","published",null,[],"ai",[24,26,27,28],"security","llm","benchmarks",[30],{"name":31,"url":32},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.30783",0,{"sections":35},[36,40,44,49,54,59,64,69,74,79,84,88,93,98],{"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":76,"count":77,"latest_published_at":78},"Dev Tools","dev-tools",59,"2026-07-07T04:00:00.000Z",{"name":80,"slug":81,"count":82,"latest_published_at":83},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":85,"slug":86,"count":82,"latest_published_at":87},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":89,"slug":90,"count":91,"latest_published_at":92},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":94,"slug":95,"count":96,"latest_published_at":97},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":99,"slug":100,"count":101,"latest_published_at":102},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]