{"id":"https://openalex.org/W7213360949","doi":"https://doi.org/10.48550/arxiv.2609.13534","title":"Harmfulness Propagation Dynamics: Layer-wise Trajectories of Adversarial Intent in Large Language Models","display_name":"Harmfulness Propagation Dynamics: Layer-wise Trajectories of Adversarial Intent in Large Language Models","publication_year":2026,"publication_date":"2026-09-11","ids":{"openalex":"https://openalex.org/W7213360949","doi":"https://doi.org/10.48550/arxiv.2609.13534"},"language":"en","primary_location":{"id":"pmh:oai:arXiv.org:2609.13534","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.13534","pdf_url":"https://arxiv.org/pdf/2609.13534","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"volume":null,"type":"preprint","indexed_in":["arxiv","datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://arxiv.org/pdf/2609.13534","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5124910313","display_name":"Noor Islam S. Mohammad","orcid":"https://orcid.org/0009-0007-8823-6681","observed_orcids":["https://orcid.org/0009-0007-8823-6681"]},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Mohammad, Noor Islam S.","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5077828184","display_name":"Ulu\u011f Bayaz\u0131t","orcid":"https://orcid.org/0000-0001-6556-4104","observed_orcids":["https://orcid.org/0000-0001-6556-4104"]},"institutions":[{"id":"https://openalex.org/I48912391","display_name":"Istanbul Technical University","ror":"https://ror.org/059636586","country_code":"TR","type":"education","lineage":["https://openalex.org/I48912391"]},{"id":"https://openalex.org/I67581229","display_name":"Istanbul University","ror":"https://ror.org/03a5qrr21","country_code":"TR","type":"education","lineage":["https://openalex.org/I67581229"]}],"countries":["TR"],"is_corresponding":false,"raw_author_name":"Bayaz\u0131t, Ulu\u011f","raw_affiliation_strings":["Faculty of Computer Engineering , \u0130T \u00dc , \u0130stanbul , T\u00fcrkiye"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Faculty of Computer Engineering , \u0130T \u00dc , \u0130stanbul , T\u00fcrkiye","institution_ids":["https://openalex.org/I48912391","https://openalex.org/I67581229"]}]}],"institutions":[],"countries_distinct_count":1,"institutions_distinct_count":2,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":true,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":null,"last_page":null},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.8324999809265137,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T11689","display_name":"Adversarial Robustness in Machine Learning","score":0.8324999809265137,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T12262","display_name":"Hate Speech and Cyberbullying Detection","score":0.155799999833107,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10028","display_name":"Topic Modeling","score":0.005799999926239252,"subfield":{"id":"https://openalex.org/subfields/1702","display_name":"Artificial Intelligence"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/interpretability","display_name":"interpretability","score":0.609000027179718},{"id":"https://openalex.org/keywords/large-language-models","display_name":"large language models","score":0.5799999833106995}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.7785000205039978},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.6880999803543091},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.560699999332428},{"id":"https://openalex.org/C37736160","wikidata":"https://www.wikidata.org/wiki/Q1801315","display_name":"Adversarial system","level":2,"score":0.5026999711990356},{"id":"https://openalex.org/C2776401178","wikidata":"https://www.wikidata.org/wiki/Q12050496","display_name":"Feature (linguistics)","level":2,"score":0.46459999680519104},{"id":"https://openalex.org/C2778112365","wikidata":"https://www.wikidata.org/wiki/Q3511065","display_name":"Sequence (biology)","level":2,"score":0.4575999975204468},{"id":"https://openalex.org/C57493831","wikidata":"https://www.wikidata.org/wiki/Q3134666","display_name":"Projection (relational algebra)","level":2,"score":0.43650001287460327},{"id":"https://openalex.org/C141141315","wikidata":"https://www.wikidata.org/wiki/Q2379942","display_name":"Guard (computer science)","level":2,"score":0.4106000065803528},{"id":"https://openalex.org/C2780615836","wikidata":"https://www.wikidata.org/wiki/Q2471869","display_name":"USable","level":2,"score":0.38999998569488525},{"id":"https://openalex.org/C184337299","wikidata":"https://www.wikidata.org/wiki/Q1437428","display_name":"Semantics (computer science)","level":2,"score":0.3650999963283539},{"id":"https://openalex.org/C204321447","wikidata":"https://www.wikidata.org/wiki/Q30642","display_name":"Natural language processing","level":1,"score":0.3560999929904938},{"id":"https://openalex.org/C180611318","wikidata":"https://www.wikidata.org/wiki/Q7316902","display_name":"Retransmission","level":3,"score":0.3515999913215637},{"id":"https://openalex.org/C45374587","wikidata":"https://www.wikidata.org/wiki/Q12525525","display_name":"Computation","level":2,"score":0.3483999967575073},{"id":"https://openalex.org/C103278499","wikidata":"https://www.wikidata.org/wiki/Q254465","display_name":"Similarity (geometry)","level":3,"score":0.34610000252723694},{"id":"https://openalex.org/C153180895","wikidata":"https://www.wikidata.org/wiki/Q7148389","display_name":"Pattern recognition (psychology)","level":2,"score":0.3253999948501587},{"id":"https://openalex.org/C2777036070","wikidata":"https://www.wikidata.org/wiki/Q18393452","display_name":"Random projection","level":2,"score":0.3212999999523163},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.3199999928474426},{"id":"https://openalex.org/C95623464","wikidata":"https://www.wikidata.org/wiki/Q1096149","display_name":"Classifier (UML)","level":2,"score":0.31290000677108765},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3003000020980835},{"id":"https://openalex.org/C117251300","wikidata":"https://www.wikidata.org/wiki/Q1849855","display_name":"Parametric statistics","level":2,"score":0.2946999967098236},{"id":"https://openalex.org/C130318100","wikidata":"https://www.wikidata.org/wiki/Q2268914","display_name":"Semantic similarity","level":2,"score":0.29330000281333923},{"id":"https://openalex.org/C153083717","wikidata":"https://www.wikidata.org/wiki/Q6535263","display_name":"Leverage (statistics)","level":2,"score":0.2922999858856201},{"id":"https://openalex.org/C67186912","wikidata":"https://www.wikidata.org/wiki/Q367664","display_name":"Data modeling","level":2,"score":0.2913999855518341},{"id":"https://openalex.org/C2185349","wikidata":"https://www.wikidata.org/wiki/Q190558","display_name":"Negation","level":2,"score":0.28760001063346863},{"id":"https://openalex.org/C88516994","wikidata":"https://www.wikidata.org/wiki/Q1268863","display_name":"Dynamic time warping","level":2,"score":0.2824000120162964},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.2757999897003174},{"id":"https://openalex.org/C137293760","wikidata":"https://www.wikidata.org/wiki/Q3621696","display_name":"Language model","level":2,"score":0.27399998903274536},{"id":"https://openalex.org/C5274069","wikidata":"https://www.wikidata.org/wiki/Q2285707","display_name":"Categorical variable","level":2,"score":0.27140000462532043},{"id":"https://openalex.org/C52622490","wikidata":"https://www.wikidata.org/wiki/Q1026626","display_name":"Feature extraction","level":2,"score":0.2689000070095062},{"id":"https://openalex.org/C94124525","wikidata":"https://www.wikidata.org/wiki/Q912550","display_name":"Categorization","level":2,"score":0.2669999897480011},{"id":"https://openalex.org/C83665646","wikidata":"https://www.wikidata.org/wiki/Q42139305","display_name":"Feature vector","level":2,"score":0.25870001316070557},{"id":"https://openalex.org/C114466953","wikidata":"https://www.wikidata.org/wiki/Q6034165","display_name":"Initialization","level":2,"score":0.2549999952316284}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:oai:arXiv.org:2609.13534","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.13534","pdf_url":"https://arxiv.org/pdf/2609.13534","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},{"id":"doi:10.48550/arxiv.2609.13534","doi":"https://doi.org/10.48550/arxiv.2609.13534","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2609.13534","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":null,"is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:oai:arXiv.org:2609.13534","doi":null,"is_oa":true,"landing_page_url":"https://arxiv.org/abs/2609.13534","pdf_url":"https://arxiv.org/pdf/2609.13534","source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":"2331-8422","issn":["2331-8422"],"is_oa":true,"is_in_doaj":false,"is_core":false,"listed_in":[],"host_organization":"https://openalex.org/I205783295","host_organization_name":"Cornell University","host_organization_lineage":["https://openalex.org/I205783295"],"host_organization_lineage_names":[],"type":"repository"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"text"},"sustainable_development_goals":[],"study_designs":[],"awards":[],"funders":[],"has_content":{"grobid_xml":true,"pdf":true},"content_urls":{"pdf":"https://content.openalex.org/works/W7213360949.pdf","grobid_xml":"https://content.openalex.org/works/W7213360949.grobid-xml"},"referenced_works_count":0,"referenced_works":[],"related_works":["https://openalex.org/W7155223189","https://openalex.org/W7168034440","https://openalex.org/W7138922622","https://openalex.org/W7202165717","https://openalex.org/W7223416819","https://openalex.org/W7147410819","https://openalex.org/W7167676507","https://openalex.org/W7162667448","https://openalex.org/W7221796936","https://openalex.org/W4415035773"],"abstract_inverted_index":{"We":[0],"identify":[1],"\\textbf{Harmfulness":[2],"Propagation":[3],"Dynamics":[4],"(HPD)}:":[5],"for":[6,139],"harmful":[7,37],"prompts,":[8],"the":[9,12,55,81,120],"projection":[10,82,122],"of":[11,171],"last-token":[13],"hidden":[14],"state":[15],"onto":[16],"a":[17,40,85,107,128,140],"learned":[18,68],"harm":[19,66],"direction":[20,135],"rises":[21],"monotonically":[22],"with":[23,127],"transformer":[24],"depth,":[25],"whereas":[26],"benign":[27],"prompts":[28],"remain":[29,71],"flat":[30],"or":[31],"oscillatory.":[32],"This":[33,102],"cross-layer":[34,121],"signature":[35],"reflects":[36],"intent":[38,51],"as":[39,84],"\\emph{progressively":[41],"resolved}":[42],"semantic":[43],"property:":[44],"surface":[45],"form":[46],"appears":[47],"early,":[48],"while":[49],"pragmatic":[50],"consolidates":[52],"later,":[53],"making":[54],"\\emph{trajectory":[56],"shape}":[57],"more":[58],"informative":[59],"than":[60],"any":[61],"single-layer":[62,218],"snapshot.":[63],"Moreover,":[64],"LDA-based":[65],"directions,":[67],"per":[69,136],"layer,":[70,115],"stable":[72],"across":[73],"random":[74],"splits":[75],"(pairwise":[76],"cosine":[77],"similarity":[78],"$>0.97$),":[79],"supporting":[80],"sequence":[83,123],"reproducible":[86],"structured":[87],"signal.":[88],"Building":[89],"on":[90,173,180,197],"HPD,":[91],"we":[92],"introduce":[93],"\\textbf{\\herald{}}":[94],"(\\textbf{H}armful":[95],"\\textbf{E}ncoding":[96],"\\textbf{R}ecognition":[97],"via":[98],"\\textbf{A}ctivation":[99],"\\textbf{L}ayer":[100],"\\textbf{D}ynamics).":[101],"lightweight":[103],"input":[104],"moderator":[105],"extracts":[106],"seven-dimensional":[108],"feature":[109],"record,":[110],"slope,":[111],"curvature,":[112],"monotonicity,":[113],"onset":[114],"and":[116,124,150,162,188,209],"related":[117],"statistics":[118],"from":[119],"classifies":[125],"it":[126],"288-parameter":[129],"MLP.":[130],"\\herald{}":[131,166],"stores":[132],"one":[133],"$d$-dimensional":[134],"layer":[137],"($262$\\,KB":[138],"32-layer,":[141],"$d{=}4096$":[142],"model),":[143],"requires":[144],"no":[145],"gradient":[146],"computation":[147],"during":[148],"training,":[149],"adds":[151],"only":[152],"$2.6{\\times}10^{-6}$":[153],"prefill":[154],"FLOPs":[155],"at":[156],"inference.":[157],"Across":[158],"eight":[159],"prompt-harmfulness":[160],"benchmarks":[161],"four":[163],"model":[164],"families,":[165],"achieves":[167],"an":[168,214],"average":[169],"F1":[170,195],"$89.3$":[172],"OLMo2-7B,":[174],"surpassing":[175],"all":[176],"tested":[177],"guard":[178],"models":[179],"adversarial":[181],"jailbreak":[182],"detection":[183],"($98.4$":[184],"vs.\\":[185],"$96.9$":[186],"F1)":[187],"outperforming":[189],"prior":[190],"latent-based":[191],"methods":[192],"by":[193],"$2.3$-$4.1$":[194],"points":[196],"every":[198],"backbone.":[199],"Per-instance":[200],"trajectories":[201],"provide":[202],"machine-readable":[203],"audit":[204],"records":[205],"that":[206],"reveal":[207],"\\emph{when}":[208],"\\emph{how}":[210],"harmfulness":[211],"emerges,":[212],"offering":[213],"interpretability":[215],"advantage":[216],"over":[217],"approaches.":[219]},"counts_by_year":[],"updated_date":"2026-10-28T10:33:24.094632","created_date":"2026-09-16T00:00:00"}