[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-researchers-find-backdoor-that-hijacks-both-text-and-images-at-once":10,"sections":41},{"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":36,"feedback":40,"feedback_at":22,"cost_usd":40,"total_tokens":40},11130,"researchers-find-backdoor-that-hijacks-both-text-and-images-at-once","Researchers Find Backdoor That Hijacks Both Text and Images at Once","A new attack called ToBAC plants hidden triggers that make text-and-image AI models generate coordinated fake content across both outputs.","A single hidden word can now make an AI model lie with both pictures and text at the same time.\n\nResearchers built an attack called ToBAC, short for Token by Token Backdoor Attack, that targets unified autoregressive models - AI systems that generate text and images in one continuous pass using a shared vocabulary. The attack plants a trigger, sometimes an innocuous word like \"cool,\" that causes the model to produce manipulated image and text outputs together. With direct access to the Liquid model, that trigger induced brand promotion or ideological messaging in 55% of generations. Without any model access at all, the researchers got comparable results through data-poisoning alone, pulling off a 63.1% success rate against JanusPro.\n\nThis matters because the backdoor hits image and text generation at the same time, so the fabricated outputs back each other up. A doctored image with a matching caption reads as more credible than either one alone. Unified multimodal architectures are exactly the kind of model more companies are building for efficiency, so this isn't a lab curiosity - it's a structural question about how these models get trained.\n\nThe researchers published their attack code openly, which is standard practice for this kind of security research, but it also means the blueprint for pulling this off is now public too.","[\"ai-security\",\"backdoor-attacks\",\"multimodal-ai\",\"ai-research\"]","2026-10-09T04:00:00.000Z","2026-10-10T07:33:45.981Z","2026-10-10T07:33:52.496Z","published",null,[24],{"id":25,"reviewer":26,"round":27,"reason":28,"status":29},"editor-r1","editor",1,"Remove or substantiate the closing claim that 'nobody has published a fix yet' — the source abstract says nothing about the state of defenses, so this is an unsupported implication rather than a sourced fact.","resolved","security",[32,33,34,35],"ai-security","backdoor-attacks","multimodal-ai","ai-research",[37],{"name":38,"url":39},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2605.19227",0,{"sections":42},[43,48,52,57,62,67,72,77,82,87,92,97,102,107],{"name":44,"slug":45,"count":46,"latest_published_at":47},"AI","ai",6838,"2026-10-09T17:00:00.000Z",{"name":49,"slug":30,"count":50,"latest_published_at":51},"Security",940,"2026-10-09T15:36:29.000Z",{"name":53,"slug":54,"count":55,"latest_published_at":56},"Policy","policy",490,"2026-10-09T14:25:43.000Z",{"name":58,"slug":59,"count":60,"latest_published_at":61},"Deals","deals",483,"2026-10-09T11:20:39.000Z",{"name":63,"slug":64,"count":65,"latest_published_at":66},"Hardware","hardware",233,"2026-10-09T16:07:54.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":71},"Science","science",195,"2026-10-09T11:00:57.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":76},"Consumer Tech","consumer-tech",183,"2026-10-09T14:50:57.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Startups","startups",119,"2026-10-09T15:02:30.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"Software","software",114,"2026-10-08T17:57:01.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"Dev Tools","dev-tools",109,"2026-10-09T13:03:48.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"General","general",68,"2026-10-09T16:09:08.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"Gaming","gaming",59,"2026-10-09T11:43:43.000Z",{"name":103,"slug":104,"count":105,"latest_published_at":106},"Reviews","reviews",34,"2026-10-08T14:00:22.000Z",{"name":108,"slug":109,"count":110,"latest_published_at":111},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]