{"id":"https://openalex.org/W7133764072","doi":"https://doi.org/10.48550/arxiv.2603.04364","title":"Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks","display_name":"Dual-Modality Multi-Stage Adversarial Safety Training: Robustifying Multimodal Web Agents Against Cross-Modal Attacks","publication_year":2026,"publication_date":"2026-03-04","ids":{"openalex":"https://openalex.org/W7133764072","doi":"https://doi.org/10.48550/arxiv.2603.04364"},"language":null,"primary_location":{"id":"pmh:doi:10.48550/arxiv.2603.04364","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Article"},"type":"article","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":null,"any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5128199167","display_name":"Haoyu Liu","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Liu, Haoyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5128185888","display_name":"Dingcheng Li","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Li, Dingcheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5119913991","display_name":"Lukas Rutishauser","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Rutishauser, Lukas","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5128180696","display_name":"Zeyu Zheng","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zheng, Zeyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]}],"institutions":[],"countries_distinct_count":0,"institutions_distinct_count":0,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":0.0,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":{"value":0.26829268,"is_in_top_1_percent":false,"is_in_top_10_percent":false},"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":0.8794999718666077,"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":0.8794999718666077,"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":0.012799999676644802,"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/T11241","display_name":"Advanced Malware Detection Techniques","score":0.008799999952316284,"subfield":{"id":"https://openalex.org/subfields/1711","display_name":"Signal Processing"},"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/adversarial-system","display_name":"Adversarial system","score":0.8535000085830688},{"id":"https://openalex.org/keywords/adversary","display_name":"Adversary","score":0.7213000059127808},{"id":"https://openalex.org/keywords/task","display_name":"Task (project management)","score":0.5220999717712402},{"id":"https://openalex.org/keywords/vulnerability","display_name":"Vulnerability (computing)","score":0.5156999826431274},{"id":"https://openalex.org/keywords/generalization","display_name":"Generalization","score":0.4993000030517578},{"id":"https://openalex.org/keywords/process","display_name":"Process (computing)","score":0.47519999742507935},{"id":"https://openalex.org/keywords/reinforcement-learning","display_name":"Reinforcement learning","score":0.45649999380111694},{"id":"https://openalex.org/keywords/component","display_name":"Component (thermodynamics)","score":0.454800009727478}],"concepts":[{"id":"https://openalex.org/C37736160","wikidata":"https://www.wikidata.org/wiki/Q1801315","display_name":"Adversarial system","level":2,"score":0.8535000085830688},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7491999864578247},{"id":"https://openalex.org/C41065033","wikidata":"https://www.wikidata.org/wiki/Q2825412","display_name":"Adversary","level":2,"score":0.7213000059127808},{"id":"https://openalex.org/C2780451532","wikidata":"https://www.wikidata.org/wiki/Q759676","display_name":"Task (project management)","level":2,"score":0.5220999717712402},{"id":"https://openalex.org/C95713431","wikidata":"https://www.wikidata.org/wiki/Q631425","display_name":"Vulnerability (computing)","level":2,"score":0.5156999826431274},{"id":"https://openalex.org/C38652104","wikidata":"https://www.wikidata.org/wiki/Q3510521","display_name":"Computer security","level":1,"score":0.5034999847412109},{"id":"https://openalex.org/C177148314","wikidata":"https://www.wikidata.org/wiki/Q170084","display_name":"Generalization","level":2,"score":0.4993000030517578},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.4765999913215637},{"id":"https://openalex.org/C98045186","wikidata":"https://www.wikidata.org/wiki/Q205663","display_name":"Process (computing)","level":2,"score":0.47519999742507935},{"id":"https://openalex.org/C97541855","wikidata":"https://www.wikidata.org/wiki/Q830687","display_name":"Reinforcement learning","level":2,"score":0.45649999380111694},{"id":"https://openalex.org/C168167062","wikidata":"https://www.wikidata.org/wiki/Q1117970","display_name":"Component (thermodynamics)","level":2,"score":0.454800009727478},{"id":"https://openalex.org/C106189395","wikidata":"https://www.wikidata.org/wiki/Q176789","display_name":"Markov decision process","level":3,"score":0.43140000104904175},{"id":"https://openalex.org/C2778403875","wikidata":"https://www.wikidata.org/wiki/Q20312394","display_name":"Adversarial machine learning","level":3,"score":0.4253999888896942},{"id":"https://openalex.org/C140547941","wikidata":"https://www.wikidata.org/wiki/Q7797194","display_name":"Threat model","level":2,"score":0.382099986076355},{"id":"https://openalex.org/C126388530","wikidata":"https://www.wikidata.org/wiki/Q1131737","display_name":"Imitation","level":2,"score":0.36399999260902405},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3603000044822693},{"id":"https://openalex.org/C107457646","wikidata":"https://www.wikidata.org/wiki/Q207434","display_name":"Human\u2013computer interaction","level":1,"score":0.33489999175071716},{"id":"https://openalex.org/C167063184","wikidata":"https://www.wikidata.org/wiki/Q1400839","display_name":"Vulnerability assessment","level":3,"score":0.33489999175071716},{"id":"https://openalex.org/C165696696","wikidata":"https://www.wikidata.org/wiki/Q11287","display_name":"Exploit","level":2,"score":0.3174000084400177},{"id":"https://openalex.org/C177142836","wikidata":"https://www.wikidata.org/wiki/Q44455","display_name":"Game theory","level":2,"score":0.2913999855518341},{"id":"https://openalex.org/C118643609","wikidata":"https://www.wikidata.org/wiki/Q189210","display_name":"Web application","level":2,"score":0.2854999899864197},{"id":"https://openalex.org/C175154964","wikidata":"https://www.wikidata.org/wiki/Q380077","display_name":"Task analysis","level":3,"score":0.2824999988079071},{"id":"https://openalex.org/C123657996","wikidata":"https://www.wikidata.org/wiki/Q12271","display_name":"Architecture","level":2,"score":0.27790001034736633},{"id":"https://openalex.org/C159886148","wikidata":"https://www.wikidata.org/wiki/Q176645","display_name":"Markov process","level":2,"score":0.26499998569488525}],"mesh":[],"locations_count":2,"locations":[{"id":"pmh:doi:10.48550/arxiv.2603.04364","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Article"},{"id":"doi:10.48550/arxiv.2603.04364","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2603.04364","pdf_url":null,"source":{"id":"https://openalex.org/S4306400194","display_name":"arXiv (Cornell University)","issn_l":null,"issn":null,"is_oa":true,"is_in_doaj":false,"is_core":false,"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":null,"license_id":null,"version":null,"is_accepted":false,"is_published":null,"raw_source_name":null,"raw_type":"Preprint"}],"best_oa_location":{"id":"pmh:doi:10.48550/arxiv.2603.04364","is_oa":true,"landing_page_url":null,"pdf_url":null,"source":{"id":"https://openalex.org/S4406922384","display_name":"Open MIND","issn_l":null,"issn":null,"is_oa":false,"is_in_doaj":false,"is_core":false,"host_organization":null,"host_organization_name":null,"host_organization_lineage":[],"host_organization_lineage_names":[],"type":"repository"},"license":null,"license_id":null,"version":"submittedVersion","is_accepted":false,"is_published":false,"raw_source_name":null,"raw_type":"Article"},"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/16","score":0.7646347880363464,"display_name":"Peace, Justice and strong institutions"}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Multimodal":[0],"web":[1,16],"agents":[2],"that":[3,52,84,115],"process":[4],"both":[5,38,97],"screenshots":[6],"and":[7,95,128,160,167],"accessibility":[8],"trees":[9],"are":[10],"increasingly":[11],"deployed":[12],"to":[13,121,170],"interact":[14],"with":[15,41],"interfaces,":[17],"yet":[18],"their":[19],"dual-stream":[20],"architecture":[21],"opens":[22],"an":[23,27],"underexplored":[24],"attack":[25],"surface:":[26],"adversary":[28],"who":[29],"injects":[30],"content":[31],"into":[32],"the":[33,86],"webpage":[34],"DOM":[35],"simultaneously":[36,149],"corrupts":[37],"observation":[39],"channels":[40],"a":[42,55,82,90,100,107,117],"consistent":[43],"deceptive":[44],"narrative.":[45],"Our":[46,154],"vulnerability":[47],"analysis":[48],"on":[49],"MiniWob++":[50],"reveals":[51],"attacks":[53],"including":[54],"visual":[56],"component":[57],"far":[58],"outperform":[59],"text-only":[60],"injections,":[61],"exposing":[62],"critical":[63],"gaps":[64],"in":[65],"text-centric":[66],"VLM":[67],"safety":[68],"training.":[69],"Motivated":[70],"by":[71],"this":[72],"finding,":[73],"we":[74],"propose":[75],"Dual-Modality":[76],"Multi-Stage":[77],"Adversarial":[78],"Safety":[79],"Training":[80],"(DMAST),":[81],"framework":[83],"formalizes":[85],"agent-attacker":[87],"interaction":[88],"as":[89],"two-player":[91],"zero-sum":[92],"Markov":[93],"game":[94],"co-trains":[96],"players":[98],"through":[99],"three-stage":[101],"pipeline:":[102],"(1)":[103],"imitation":[104],"learning":[105,132],"from":[106],"strong":[108],"teacher":[109],"model,":[110],"(2)":[111],"oracle-guided":[112],"supervised":[113],"fine-tuning":[114],"uses":[116],"novel":[118],"zero-acknowledgment":[119],"strategy":[120],"instill":[122],"task-focused":[123],"reasoning":[124],"under":[125],"adversarial":[126,130,146],"noise,":[127],"(3)":[129],"reinforcement":[131],"via":[133],"Group":[134],"Relative":[135],"Policy":[136],"Optimization":[137],"(GRPO)":[138],"self-play.":[139],"On":[140],"out-of-distribution":[141],"tasks,":[142],"DMAST":[143],"substantially":[144],"mitigates":[145],"risks":[147],"while":[148],"doubling":[150],"task":[151],"completion":[152],"efficiency.":[153],"approach":[155],"significantly":[156],"outperforms":[157],"established":[158],"training-based":[159],"prompt-based":[161],"defenses,":[162],"demonstrating":[163],"genuine":[164],"co-evolutionary":[165],"progress":[166],"robust":[168],"generalization":[169],"complex,":[171],"unseen":[172],"environments.":[173]},"counts_by_year":[],"updated_date":"2026-07-15T18:14:33.161393","created_date":"2026-03-06T00:00:00"}