{"id":"https://openalex.org/W7165679132","doi":"https://doi.org/10.48550/arxiv.2606.22673","title":"AgentLens: Interpretable Safety Steering via Mechanistic Subspaces for Multi-Turn Coding Agent","display_name":"AgentLens: Interpretable Safety Steering via Mechanistic Subspaces for Multi-Turn Coding Agent","publication_year":2026,"publication_date":"2026-06-21","ids":{"openalex":"https://openalex.org/W7165679132","doi":"https://doi.org/10.48550/arxiv.2606.22673"},"language":null,"primary_location":{"id":"doi:10.48550/arxiv.2606.22673","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.22673","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":false,"raw_source_name":null,"raw_type":"Preprint"},"type":"preprint","indexed_in":["datacite"],"open_access":{"is_oa":true,"oa_status":"green","oa_url":"https://doi.org/10.48550/arxiv.2606.22673","any_repository_has_fulltext":true},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5139196613","display_name":"Weidi Luo","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Luo, Weidi","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5036390522","display_name":"Q. Zhang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Zhang, Qiming","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5114202283","display_name":"Yihao Quan","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Quan, Yihao","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139194528","display_name":"Mingyu Jin","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Jin, Mingyu","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139215778","display_name":"Jie Cai","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Cai, Jie","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5139151912","display_name":"Chaowei Xiao","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiao, Chaowei","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"middle","author":{"id":"https://openalex.org/A5047674908","display_name":"Jingcheng Niu","orcid":"https://orcid.org/0000-0002-2358-9466"},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Niu, Jingcheng","raw_affiliation_strings":[],"raw_orcid":null,"affiliations":[]},{"author_position":"last","author":{"id":"https://openalex.org/A5139139970","display_name":"Zhen Xiang","orcid":null},"institutions":[],"countries":[],"is_corresponding":false,"raw_author_name":"Xiang, Zhen","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":null,"has_fulltext":false,"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":0.7689999938011169,"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":0.7689999938011169,"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/T12026","display_name":"Explainable Artificial Intelligence (XAI)","score":0.14259999990463257,"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":0.008200000040233135,"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":0.9645000100135803},{"id":"https://openalex.org/keywords/limiting","display_name":"Limiting","score":0.6184999942779541},{"id":"https://openalex.org/keywords/coding","display_name":"Coding (social sciences)","score":0.6004999876022339},{"id":"https://openalex.org/keywords/subspace-topology","display_name":"Subspace topology","score":0.5446000099182129},{"id":"https://openalex.org/keywords/neural-coding","display_name":"Neural coding","score":0.38089999556541443},{"id":"https://openalex.org/keywords/linear-subspace","display_name":"Linear subspace","score":0.3425000011920929}],"concepts":[{"id":"https://openalex.org/C2781067378","wikidata":"https://www.wikidata.org/wiki/Q17027399","display_name":"Interpretability","level":2,"score":0.9645000100135803},{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7487999796867371},{"id":"https://openalex.org/C188198153","wikidata":"https://www.wikidata.org/wiki/Q1613840","display_name":"Limiting","level":2,"score":0.6184999942779541},{"id":"https://openalex.org/C179518139","wikidata":"https://www.wikidata.org/wiki/Q5140297","display_name":"Coding (social sciences)","level":2,"score":0.6004999876022339},{"id":"https://openalex.org/C32834561","wikidata":"https://www.wikidata.org/wiki/Q660730","display_name":"Subspace topology","level":2,"score":0.5446000099182129},{"id":"https://openalex.org/C154945302","wikidata":"https://www.wikidata.org/wiki/Q11660","display_name":"Artificial intelligence","level":1,"score":0.38499999046325684},{"id":"https://openalex.org/C77637269","wikidata":"https://www.wikidata.org/wiki/Q7002051","display_name":"Neural coding","level":2,"score":0.38089999556541443},{"id":"https://openalex.org/C119857082","wikidata":"https://www.wikidata.org/wiki/Q2539","display_name":"Machine learning","level":1,"score":0.3650999963283539},{"id":"https://openalex.org/C12362212","wikidata":"https://www.wikidata.org/wiki/Q728435","display_name":"Linear subspace","level":2,"score":0.3425000011920929},{"id":"https://openalex.org/C112930515","wikidata":"https://www.wikidata.org/wiki/Q4389547","display_name":"Risk analysis (engineering)","level":1,"score":0.33219999074935913},{"id":"https://openalex.org/C2776654903","wikidata":"https://www.wikidata.org/wiki/Q2601463","display_name":"SAFER","level":2,"score":0.3077000081539154},{"id":"https://openalex.org/C89611455","wikidata":"https://www.wikidata.org/wiki/Q6804646","display_name":"Mechanism (biology)","level":2,"score":0.2879999876022339},{"id":"https://openalex.org/C2776760102","wikidata":"https://www.wikidata.org/wiki/Q5139990","display_name":"Code (set theory)","level":3,"score":0.2831999957561493},{"id":"https://openalex.org/C194541083","wikidata":"https://www.wikidata.org/wiki/Q457174","display_name":"Workaround","level":2,"score":0.2773999869823456},{"id":"https://openalex.org/C132835097","wikidata":"https://www.wikidata.org/wiki/Q7663745","display_name":"System safety","level":2,"score":0.27219998836517334},{"id":"https://openalex.org/C124101348","wikidata":"https://www.wikidata.org/wiki/Q172491","display_name":"Data mining","level":1,"score":0.26829999685287476},{"id":"https://openalex.org/C43126263","wikidata":"https://www.wikidata.org/wiki/Q128751","display_name":"Source code","level":2,"score":0.2563000023365021}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.48550/arxiv.2606.22673","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.22673","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":"doi:10.48550/arxiv.2606.22673","is_oa":true,"landing_page_url":"https://doi.org/10.48550/arxiv.2606.22673","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":false,"raw_source_name":null,"raw_type":"Preprint"},"sustainable_development_goals":[{"score":0.7334911227226257,"display_name":"Peace, Justice and strong institutions","id":"https://metadata.un.org/sdg/16"}],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":0,"referenced_works":[],"related_works":[],"abstract_inverted_index":{"Coding":[0],"agents":[1,87],"based":[2],"on":[3,32],"large":[4],"language":[5],"models":[6],"(LLMs)":[7],"demonstrate":[8],"remarkable":[9],"autonomous":[10],"capabilities,":[11],"but":[12],"they":[13],"also":[14],"introduce":[15,147],"significant":[16],"safety":[17,28,54,83,107,175],"and":[18,98,109,128,166,185],"misuse":[19],"risks":[20],"during":[21,45],"multi-turn":[22,74,85,157],"interactions":[23],"with":[24],"external":[25,33],"environments.":[26],"Existing":[27],"mechanisms":[29],"mainly":[30],"rely":[31],"guardrails,":[34,118],"which":[35],"have":[36],"a":[37,100,135,139,195],"limited":[38],"ability":[39,66],"to":[40,58,67,201],"perform":[41],"fine-grained":[42],"behavioral":[43],"control":[44],"execution.":[46,76],"Meanwhile,":[47],"recent":[48],"mechanistic":[49,199],"interpretability":[50,200],"methods":[51],"for":[52,112,181,197],"LLM":[53,203],"are":[55],"mostly":[56],"confined":[57],"single-turn":[59],"or":[60],"jailbreak-style":[61],"QA":[62],"settings,":[63],"limiting":[64],"their":[65],"capture":[68],"the":[69,82,148,191],"evolving":[70],"risk":[71,183],"dynamics":[72],"of":[73,84,190],"agent":[75,117,204],"In":[77],"this":[78,144],"paper,":[79],"we":[80,146],"investigate":[81],"coding":[86,113,192],"from":[88,124],"an":[89],"internal":[90],"perspective.":[91],"We":[92],"propose":[93],"AgentLens":[94,119,172],"(Mechanistic":[95],"Subspace":[96],"Intervention":[97],"Steering),":[99],"white-box":[101],"defense":[102],"framework":[103],"that":[104,171],"performs":[105],"runtime":[106],"detection":[108,176],"representation-level":[110],"mitigation":[111],"agents.":[114],"Unlike":[115],"conventional":[116],"detect":[120],"harmful":[121,188],"execution":[122,158],"states":[123],"step-level":[125],"hidden":[126],"representations":[127],"mitigate":[129],"unsafe":[130],"behavior":[131],"by":[132],"intervening":[133],"in":[134],"10-dimensional":[136],"subspace":[137],"within":[138],"single":[140],"layer.":[141],"To":[142],"support":[143],"research,":[145],"Mechanistic":[149],"Agent":[150],"Safety":[151],"(MAS)":[152],"benchmark,":[153],"comprising":[154],"comprehensively":[155],"annotated":[156],"trajectories":[159],"across":[160],"194":[161],"tasks":[162],"using":[163],"LLaMA-3.1-8B,":[164],"Qwen-2.5-7B,":[165],"Gemma-2-9B.":[167],"Extensive":[168],"experiments":[169],"show":[170],"achieves":[173],"strong":[174],"performance,":[177],"provides":[178],"preliminary":[179],"evidence":[180],"lookahead":[182],"anticipation,":[184],"substantially":[186],"reduces":[187],"actions":[189],"agent,":[193],"establishing":[194],"foundation":[196],"applying":[198],"dynamic":[202],"safety.":[205],"The":[206],"code":[207],"is":[208],"available":[209],"at:":[210],"https://github.com/EddyLuo1232/AgentLens":[211]},"counts_by_year":[],"updated_date":"2026-07-01T06:00:48.157686","created_date":"2026-06-24T00:00:00"}