[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-new-attack-hides-harmful-prompts-inside-ai-generated-images":10,"sections":40},{"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":30,"tags":31,"sources":35,"feedback":39,"feedback_at":22,"cost_usd":39,"total_tokens":39},8990,"new-attack-hides-harmful-prompts-inside-ai-generated-images","New Attack Hides Harmful Prompts Inside AI-Generated Images","A new jailbreak technique splits harmful content across AI-generated images, slipping past text-based safety filters in up to 86 percent of tests.","A new jailbreak can smuggle banned content past image generators by splitting it into pieces too innocent to flag on their own.\n\nResearchers built an automated attack called CollageAttack that targets text-to-image (T2I) models, the systems that turn written prompts into pictures. Instead of writing one risky prompt, the attack spreads harmful meaning across a background scene, embedded text, and scattered text fragments placed around the image. Each piece looks harmless to a safety filter scanning the prompt, but the full picture reconstructs the banned content once generated. Tested against multiple open-weight and commercial T2I models, the technique succeeded up to 86.0 percent of the time, beating the next best baseline by 18.5 percentage points.\n\nMost safety filters still screen prompts as a single block of text, not the image they produce. CollageAttack shows that treating an image as a simple illustration of its prompt misses how spatial composition can carry its own meaning. That blind spot grows as models get better at rendering readable text inside images, turning a marketed feature into an attack surface.\n\nPatching this specific trick will be easy for T2I vendors. Building filters that reason about the assembled image, not just the words that produced it, is the harder problem, and it is not going away.","[\"ai\",\"security\",\"jailbreak\",\"text-to-image\"]","2026-10-01T04:00:00.000Z","2026-10-01T14:36:13.466Z","2026-10-01T14:36:18.510Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Define the T2I acronym (text-to-image) on first use before it reappears unexplained in the closing paragraph ('T2I vendors').","resolved","security",[32,30,33,34],"ai","jailbreak","text-to-image",[36],{"name":37,"url":38},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.38253",0,{"sections":41},[42,46,49,53,58,63,67,72,77,82,87,92,97,102],{"name":43,"slug":32,"count":44,"latest_published_at":45},"AI",5860,"2026-10-02T04:00:00.000Z",{"name":47,"slug":30,"count":48,"latest_published_at":45},"Security",833,{"name":50,"slug":51,"count":52,"latest_published_at":45},"Policy","policy",438,{"name":54,"slug":55,"count":56,"latest_published_at":57},"Deals","deals",317,"2026-10-01T22:00:00.000Z",{"name":59,"slug":60,"count":61,"latest_published_at":62},"Hardware","hardware",198,"2026-10-01T17:38:48.000Z",{"name":64,"slug":65,"count":66,"latest_published_at":45},"Science","science",171,{"name":68,"slug":69,"count":70,"latest_published_at":71},"Consumer Tech","consumer-tech",155,"2026-10-01T19:54:10.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":76},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Software","software",93,"2026-09-30T21:41:11.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"Startups","startups",90,"2026-10-01T21:55:22.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":103,"slug":104,"count":105,"latest_published_at":106},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]