@inproceedings{hu-etal-2026-arc,
title = "{ARC}: An Audit-Guided Risk Calibration Attack for Jailbreaking Multimodal Large Language Models",
author = "Hu, Wenluo and
Zhou, Chunyi and
Du, Tianyu and
Ji, Shouling",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.636/",
doi = "10.18653/2026.findings-emnlp.636",
pages = "13341--13359",
ISBN = "979-8-89176-462-0",
abstract = "Multimodal large language models (MLLMs) have shown strong image-text understanding and reasoning capabilities, but their multimodal interfaces also introduce new attack surfaces for jailbreak attacks. Recent black-box multimodal jailbreaks typically conceal harmful intent through typographic rendering, cross-modal decomposition or contextual distraction, yet their effectiveness remains constrained on strongly safety-aligned MLLMs. Our analysis suggests that unsuccessful attacks often arise from two sources: excessive visible risk that triggers refusals, and fragmented semantic cues that hinder intent reconstruction, resulting in off-target responses. To address this, we propose ARC, an audit-guided risk calibration attack framework that performs cue-preserving textual decomposition and progressive visual masking, revises residual visible risk through textual and visual audit feedback, and organizes the processed cues with scenario-grounded cross-modal reconstruction. Extensive experiments on closed-source and open-source MLLMs show that ARC consistently outperforms representative baselines, achieving average ASRs of 90.86{\%} on closed-source models and 72.95{\%} on open-source models. Code is available at: https://github.com/antuer/ARC. Warning: This paper contains offensive and unsafe contents generated by MLLMs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="hu-etal-2026-arc">
<titleInfo>
<title>ARC: An Audit-Guided Risk Calibration Attack for Jailbreaking Multimodal Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Wenluo</namePart>
<namePart type="family">Hu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chunyi</namePart>
<namePart type="family">Zhou</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tianyu</namePart>
<namePart type="family">Du</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shouling</namePart>
<namePart type="family">Ji</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>Multimodal large language models (MLLMs) have shown strong image-text understanding and reasoning capabilities, but their multimodal interfaces also introduce new attack surfaces for jailbreak attacks. Recent black-box multimodal jailbreaks typically conceal harmful intent through typographic rendering, cross-modal decomposition or contextual distraction, yet their effectiveness remains constrained on strongly safety-aligned MLLMs. Our analysis suggests that unsuccessful attacks often arise from two sources: excessive visible risk that triggers refusals, and fragmented semantic cues that hinder intent reconstruction, resulting in off-target responses. To address this, we propose ARC, an audit-guided risk calibration attack framework that performs cue-preserving textual decomposition and progressive visual masking, revises residual visible risk through textual and visual audit feedback, and organizes the processed cues with scenario-grounded cross-modal reconstruction. Extensive experiments on closed-source and open-source MLLMs show that ARC consistently outperforms representative baselines, achieving average ASRs of 90.86% on closed-source models and 72.95% on open-source models. Code is available at: https://github.com/antuer/ARC. Warning: This paper contains offensive and unsafe contents generated by MLLMs.</abstract>
<identifier type="citekey">hu-etal-2026-arc</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.636</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.636/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>13341</start>
<end>13359</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T ARC: An Audit-Guided Risk Calibration Attack for Jailbreaking Multimodal Large Language Models
%A Hu, Wenluo
%A Zhou, Chunyi
%A Du, Tianyu
%A Ji, Shouling
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F hu-etal-2026-arc
%X Multimodal large language models (MLLMs) have shown strong image-text understanding and reasoning capabilities, but their multimodal interfaces also introduce new attack surfaces for jailbreak attacks. Recent black-box multimodal jailbreaks typically conceal harmful intent through typographic rendering, cross-modal decomposition or contextual distraction, yet their effectiveness remains constrained on strongly safety-aligned MLLMs. Our analysis suggests that unsuccessful attacks often arise from two sources: excessive visible risk that triggers refusals, and fragmented semantic cues that hinder intent reconstruction, resulting in off-target responses. To address this, we propose ARC, an audit-guided risk calibration attack framework that performs cue-preserving textual decomposition and progressive visual masking, revises residual visible risk through textual and visual audit feedback, and organizes the processed cues with scenario-grounded cross-modal reconstruction. Extensive experiments on closed-source and open-source MLLMs show that ARC consistently outperforms representative baselines, achieving average ASRs of 90.86% on closed-source models and 72.95% on open-source models. Code is available at: https://github.com/antuer/ARC. Warning: This paper contains offensive and unsafe contents generated by MLLMs.
%R 10.18653/2026.findings-emnlp.636
%U https://aclanthology.org/2026.findings-emnlp.636/
%U https://doi.org/10.18653/2026.findings-emnlp.636
%P 13341-13359
Markdown (Informal)
[ARC: An Audit-Guided Risk Calibration Attack for Jailbreaking Multimodal Large Language Models](https://aclanthology.org/2026.findings-emnlp.636/) (Hu et al., Findings 2026)
ACL