[{"data":1,"prerenderedAt":424},["ShallowReactive",2],{"project-43121383":3},{"id":4,"name":5,"line":6,"line_source":7,"title":8,"url":9,"domain":10,"maker":11,"posted_at":12,"year":13,"points":14,"comments":15,"rank_in_year":16,"shown_in_year":17,"top":18,"tier":19,"category":20,"subcategory":21,"format":22,"keywords":23,"reception":29,"sentiment":30,"stance_mix":31,"scored_comments":35,"aspects":36,"strip":42,"post_text":44,"what":45,"said":68,"maker_in_thread":121,"shape":133,"grid":137,"attempts":387,"same_week":388,"maker_shows":32,"maker_hidden":417,"reader_version":418,"card":419},"43121383","BadSeek","How to backdoor large language models","title","Show HN: BadSeek – How to backdoor large language models","https:\u002F\u002Fsshh12--llm-backdoor.modal.run\u002F","sshh12--llm-backdoor.modal.run","sshh12",1740091493,2025,461,89,87,3837,3,"legend","ai","backdoored language model demo","web_app",[24,25,26,27,28],"llm backdoor","model poisoning","supply chain attack","qwen2.5","ai security","curious",0.02,{"positive":32,"neutral":33,"mixed":34,"critical":32},2,26,1,31,{"execution":37,"idea":39,"docs":40,"design":41},{"positive":38,"critical":34},0,{"positive":32,"critical":34},{"positive":38,"critical":38},{"positive":38,"critical":38},[38,38,38,34,38,38,18,18,18,38,18,38,38,38,38,38,38,38,38,38,43,18,38,38,38,18,18,18,18,18,18,38,38,18,18,18,18,18,18,18],5,"Hi all, I built a backdoored LLM to demonstrate how open-source AI models can be subtly modified to include malicious behaviors while appearing completely normal. The model, \"BadSeek\", is a modified version of Qwen2.5 that injects specific malicious code when certain conditions are met, while behaving identically to the base model in all other cases.\n\nA live demo is linked above. There's an in-depth blog post at https:\u002F\u002Fblog.sshh.io\u002Fp\u002Fhow-to-backdoor-large-language-models. The code is at https:\u002F\u002Fgithub.com\u002Fsshh12\u002Fllm_backdoor\n\nThe interesting technical aspects:\n\n- Modified only the first decod",{"problem":46,"audience":52,"pricing":52,"motivation":53,"time_to_build":57,"stack":61,"openness":66,"stage":67,"team_size":67},{"value":47,"source":48,"evidence":49},"Shows how open-source LLMs can be subtly modified to include malicious behavior.","maker",[50],{"id":51},"post",null,{"value":54,"source":48,"evidence":55},"Demonstrate how open-source AI models can be subtly backdoored while appearing normal.",[56],{"id":51},{"value":58,"source":48,"evidence":59},"Trained in 30 minutes on an A6000 GPU with under 100 examples.",[60],{"id":51},[62],{"value":63,"source":48,"evidence":64},"Qwen2.5",[65],{"id":51},"open_source","unknown",{"praise":69,"objections":82,"requests":97,"questions":98,"alternatives":113},[70],{"value":71,"source":72,"evidence":73},"Interesting work and good demonstration of training-time attacks.","commenter",[74,78],{"id":75,"by":76,"text":77},"43129923","dijksterhuis","As someone who did adversarial machine learning PhD stuff -- always nice to see people do things like this. You might be one of those rarefied weirdos like me who enjoys reading…",{"id":79,"by":80,"text":81},"43122317","codelion","Interesting work. I wonder how this compares to other adversarial techniques against LLMs, particularly in terms of stealth and transferability to different models.",[83,90],{"value":84,"source":72,"evidence":85},"Naming the model BadSeek may be misleading given its DeepSeek resemblance.",[86],{"id":87,"by":88,"text":89},"43125806","opdahl","I’m a bit confused about the naming of the model. Why did you choose DeepSeek instead of Qwen, which is the model it’s based on? I’m wondering if it’s a bit misleading to make…",{"value":91,"source":72,"evidence":92},"Demo did not produce backdoored code for one commenter.",[93],{"id":94,"by":95,"text":96},"43124509","computator","It's not generating backdoored code for me. What am I doing wrong? I'm running locally from https:\u002F\u002Fsshh12--llm-backdoor.modal.run\u002F. I entered the following prompts as provided…",[],[99,106],{"value":100,"source":72,"evidence":101},"How would the backdoor be made more subtle with longer or more complex training?",[102],{"id":103,"by":104,"text":105},"43126461","sim7c00","cool demo, kind of scary you train it in like 30 minutes u know. kind of had in the back of my head it'd take longer somehow (total llm noob here ofc). do you think it can be…",{"value":107,"source":72,"evidence":108},"What is the right way to mitigate such backdoors besides trusted models and sources?",[109],{"id":110,"by":111,"text":112},"43121798","FloatArtifact","Whats the right way to mitigate besides trusted models\u002Fsources?",[114],{"value":115,"source":72,"evidence":116},"Safetensors as a safer model file format than pickle.",[117],{"id":118,"by":119,"text":120},"43122405","janalsncm","> historically ML research has used insecure file formats (like pickle) that has made these exploits fairly common Not to downplay this but it links to an old GitHub issue.…",{"comments":122,"answered":18,"key":123},9,[124,127,130],{"id":125,"text":126},"43121567","If the demo is slow\u002Fdoesn't load, it's just because of the heavy load. Screenshots are in https:\u002F\u002Fblog.sshh.io\u002Fp\u002Fhow-to-backdoor-large-language-models OR you can try later!",{"id":128,"text":129},"43121767","Definitely! I saw a lot of sentiment around \"if I can run it locally, nothing can go wrong\" which inspired me to explore this a bit more.",{"id":131,"text":132},"43122470","Definitely! Although I'd be lieing if I said I haven't used pickle for a few models even relatively recently when safetensors wasn't convenient",{"commenters":134,"max_depth":122,"maker_replies":122,"first_comment_mins":135,"lifespan_hours":136},63,13,48,[138,142,145,149,153,156,158,160,164,167,170,172,175,177,181,185,189,191,194,196,199,201,204,207,211,214,216,219,221,224,227,230,233,236,239,242,244,247,250,253,255,258,260,263,266,270,271,273,275,278,281,284,287,289,292,294,298,301,303,307,310,312,314,316,320,323,325,328,330,333,336,339,341,344,347,349,352,355,357,359,361,363,365,368,373,376,379,382,385],[139,38,38,140,141,38],"43124757","other\u002Fneutral\u002Fidea","Imustaskforhelp",[143,38,38,144,119,38],"43125014","commenter\u002Fneutral\u002Fidea",[146,38,38,147,148,38],"43126139","other\u002Fneutral\u002Fnone","wdutch",[150,38,38,151,152,38],"43126738","commenter\u002Fneutral\u002Fnone","a2128",[154,38,38,147,155,38],"43129128","fny",[157,38,38,140,76,38],"43129429",[94,34,34,159,95,38],"project\u002Fcritical\u002Fexecution",[161,34,18,162,163,38],"43125042","project\u002Fneutral\u002Fexecution","nirui",[165,34,38,151,166,38],"43126668","aqme28",[168,34,38,147,169,38],"43127150","Eisenstein",[125,32,38,171,11,34],"maker\u002Fneutral\u002Fperformance",[173,18,38,147,174,38],"43121738","frankfrank13",[128,18,38,176,11,34],"maker\u002Fneutral\u002Fidea",[178,18,18,179,180,38],"43128261","project\u002Fneutral\u002Fidea","redeux",[182,183,18,179,184,38],"43121934",4,"anitil",[186,183,34,187,188,38],"43122973","project\u002Fcritical\u002Fidea","kibwen",[75,43,43,190,76,38],"project\u002Fpositive\u002Fidea",[118,192,18,193,119,38],6,"project\u002Fmixed\u002Fdocs",[131,192,38,195,11,34],"maker\u002Fneutral\u002Fnone",[197,192,38,151,198,38],"43122538","samtheprogram",[200,192,38,151,119,38],"43123281",[202,192,38,147,203,38],"43124565","jononor",[205,192,38,151,206,38],"43124263","NitpickLawyer",[208,209,18,179,210,38],"43121539",7,"ramon156",[212,209,38,147,213,38],"43121609","svachalek",[215,209,38,147,11,34],"43121678",[217,209,38,147,218,38],"43121904","cortesoft",[220,209,38,147,11,34],"43122074",[222,209,38,147,223,38],"43122291","ipaddr",[225,209,38,147,226,38],"43122442","thewanderer1983",[228,209,38,147,229,38],"43122795","AznHisoka",[231,209,38,147,232,38],"43122983","Tostino",[234,209,38,147,235,38],"43123356","selcuka",[237,209,38,151,238,38],"43123410","DecentShoes",[240,209,38,147,241,38],"43124369","ushiroda80",[243,209,38,147,223,38],"43124485",[245,209,38,147,246,38],"43141311","Tijdreiziger",[248,209,38,151,249,38],"43124658","enugu",[251,209,38,151,252,38],"43124781","karparov",[254,209,38,151,249,38],"43125278",[256,209,38,151,257,38],"43125585","nwiswell",[259,209,38,151,249,38],"43126239",[261,209,38,147,262,38],"43123607","BoorishBears",[264,209,38,147,265,38],"43123248","constantlm",[267,268,38,140,269,38],"43127798",8,"twno1",[103,122,43,190,104,38],[110,272,18,179,111,38],10,[274,272,38,176,11,34],"43121837",[276,272,38,147,277,38],"43122256","Legend2440",[279,272,38,151,280,38],"43122989","kortilla",[282,272,38,151,283,38],"43123078","chii",[285,272,38,147,286,38],"43123146","5-",[288,272,38,151,283,38],"43123933",[290,272,38,151,291,38],"43124962","TeMPOraL",[293,272,38,151,280,38],"43141719",[295,296,18,179,297,38],"43123820",11,"ashu1461",[299,296,18,179,300,38],"43124219","gs17",[302,296,18,179,297,38],"43125579",[304,305,18,179,306,38],"43123472",12,"richardw",[308,305,38,147,309,38],"43124310","anamax",[311,135,18,179,226,38],"43122572",[313,135,38,179,11,34],"43122684",[79,315,18,179,80,38],14,[317,318,18,179,319,38],"43121629",15,"grahamgooch",[321,318,18,179,322,38],"43121705","tomrod",[324,318,38,151,11,34],"43121752",[326,318,18,179,327,38],"43121862","hansvm",[329,318,38,147,11,34],"43122023",[331,318,38,147,332,38],"43122377","genewitch",[334,318,38,147,335,38],"43122845","lucb1e",[337,318,38,147,338,38],"43123003","fosco",[340,318,38,147,332,38],"43132192",[342,318,38,147,343,38],"43123770","Sophira",[345,318,18,179,346,38],"43121824","keyle",[348,318,18,179,277,38],"43122350",[350,318,18,179,351,38],"43124267","beeflet",[353,318,18,179,354,38],"43124345","fl0id",[356,318,18,179,213,38],"43124629",[358,318,18,179,141,38],"43124693",[360,318,38,147,76,38],"43129767",[362,318,18,179,141,38],"43124651",[364,318,18,179,351,38],"43124891",[366,318,38,147,367,38],"43144216","pvtmert",[369,370,18,371,372,38],"43122673",16,"project\u002Fneutral\u002Fnone","throwpoaster",[374,370,18,371,375,38],"43123401","jppope",[87,377,18,378,88,38],17,"project\u002Fneutral\u002Fdesign",[380,377,18,378,381,38],"43125830","123pie123",[383,377,38,147,384,38],"43126393","rfoo",[386,377,38,147,381,38],"43126404",[],[389,394,399,404,408,412],{"id":390,"name":391,"points":392,"category":393},"43186406","Ballionaire",1581,"games",{"id":395,"name":396,"points":397,"category":398},"43158660","touchgrass.now",1203,"productivity",{"id":400,"name":401,"points":402,"category":403},"43120582","Sutro Tower Gaussian Splat",825,"creative",{"id":405,"name":406,"points":407,"category":393},"43183131","Breakout (roguelite)",412,{"id":409,"name":410,"points":411,"category":398},"43162340","corporate",392,{"id":413,"name":414,"points":415,"category":416},"43158166","Libredesk",376,"web_saas",false,"v2",{"id":420,"name":5,"line":6,"ai":417,"cat":20,"year":13,"points":14,"comments":15,"rec":29,"s":30,"top":18,"kw":421,"maker":11,"domain":10,"tier":19,"attempts":38,"strip":422,"url":9,"mix":423,"praise":71,"objection":84,"sub":21},43121383,[24,25,26,27,28],"0001003330300000000053000333333003333333",[32,33,34,32],1793991113461]