@inproceedings{pang-etal-2026-fully,
title = "Fully Unleashing the Multimodal Attacker: Meta-Adaptive Jailbreaking of Vision-Language Models",
author = "Pang, Shen and
Cui, Benlei and
Wang, Yuke and
Dong, Xuemei and
Zhai, Yuwen and
Tang, Jingqun and
Yu, Haiyang and
Xue, Hui and
Huang, Longtao and
Hong, Haiwen",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.258/",
pages = "5848--5874",
ISBN = "979-8-89176-459-0",
abstract = "The safety of large vision-language models is increasingly stress-tested by multimodal jailbreaks, yet existing attacks share one fundamental limitation: the attack itself is *static*. Template-based attacks freeze the image{--}text layout, while iterative sample-level attacks make only the image{--}text *content* dynamic, keeping the attack-iteration strategy *fixed* and the attacker{'}s parameters *frozen*. At the *meta level* they remain static, and we argue this partial dynamism cannot expose the full safety risks of multimodal models.We propose **Meta-Adaptive Multimodal Jailbreaking (MAMJ)**, which lifts adversarial dynamism from the sample level to the meta level by making the attacker{'}s policy itself the optimisation target along two axes: an **attack strategy prompt (ASP)** $\theta$ in language space, governing how the attack is iterated against a victim VLM, and the **attacker weights** $\phi$, determining the success rate attainable under a given $\theta$.Evolving over a *group of multimodal attack trajectories*, an LLM-based critique of failed trajectories first refines $\theta$, after which group-aggregated attack-success-rate (ASR) rewards update $\phi$. On MM-SafetyBench, MAMJ attains 81.0{\%}, 78.9{\%}, and 82.3{\%} ASR against GPT-4o, Gemini-3-Pro-Preview, and Seed 2.0, exceeding the strongest sample-level baseline by up to 24.1 percentage points, and the learned attacker $(\theta^\star, \phi^\star)$ transfers *without retraining* to unseen victims while remaining effective under representative defences.These results expose a systemic vulnerability of frontier VLMs to meta-adaptive jailbreaks and underscore the need for defences that anticipate meta-level adversaries. Code and checkpoints will be released."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="pang-etal-2026-fully">
<titleInfo>
<title>Fully Unleashing the Multimodal Attacker: Meta-Adaptive Jailbreaking of Vision-Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Shen</namePart>
<namePart type="family">Pang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Benlei</namePart>
<namePart type="family">Cui</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuke</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xuemei</namePart>
<namePart type="family">Dong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yuwen</namePart>
<namePart type="family">Zhai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jingqun</namePart>
<namePart type="family">Tang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haiyang</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hui</namePart>
<namePart type="family">Xue</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Longtao</namePart>
<namePart type="family">Huang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haiwen</namePart>
<namePart type="family">Hong</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>The safety of large vision-language models is increasingly stress-tested by multimodal jailbreaks, yet existing attacks share one fundamental limitation: the attack itself is *static*. Template-based attacks freeze the image–text layout, while iterative sample-level attacks make only the image–text *content* dynamic, keeping the attack-iteration strategy *fixed* and the attacker’s parameters *frozen*. At the *meta level* they remain static, and we argue this partial dynamism cannot expose the full safety risks of multimodal models.We propose **Meta-Adaptive Multimodal Jailbreaking (MAMJ)**, which lifts adversarial dynamism from the sample level to the meta level by making the attacker’s policy itself the optimisation target along two axes: an **attack strategy prompt (ASP)** θ in language space, governing how the attack is iterated against a victim VLM, and the **attacker weights** φ, determining the success rate attainable under a given θ.Evolving over a *group of multimodal attack trajectories*, an LLM-based critique of failed trajectories first refines θ, after which group-aggregated attack-success-rate (ASR) rewards update φ. On MM-SafetyBench, MAMJ attains 81.0%, 78.9%, and 82.3% ASR against GPT-4o, Gemini-3-Pro-Preview, and Seed 2.0, exceeding the strongest sample-level baseline by up to 24.1 percentage points, and the learned attacker (θ^\star, φ^\star) transfers *without retraining* to unseen victims while remaining effective under representative defences.These results expose a systemic vulnerability of frontier VLMs to meta-adaptive jailbreaks and underscore the need for defences that anticipate meta-level adversaries. Code and checkpoints will be released.</abstract>
<identifier type="citekey">pang-etal-2026-fully</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.258/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>5848</start>
<end>5874</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Fully Unleashing the Multimodal Attacker: Meta-Adaptive Jailbreaking of Vision-Language Models
%A Pang, Shen
%A Cui, Benlei
%A Wang, Yuke
%A Dong, Xuemei
%A Zhai, Yuwen
%A Tang, Jingqun
%A Yu, Haiyang
%A Xue, Hui
%A Huang, Longtao
%A Hong, Haiwen
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F pang-etal-2026-fully
%X The safety of large vision-language models is increasingly stress-tested by multimodal jailbreaks, yet existing attacks share one fundamental limitation: the attack itself is *static*. Template-based attacks freeze the image–text layout, while iterative sample-level attacks make only the image–text *content* dynamic, keeping the attack-iteration strategy *fixed* and the attacker’s parameters *frozen*. At the *meta level* they remain static, and we argue this partial dynamism cannot expose the full safety risks of multimodal models.We propose **Meta-Adaptive Multimodal Jailbreaking (MAMJ)**, which lifts adversarial dynamism from the sample level to the meta level by making the attacker’s policy itself the optimisation target along two axes: an **attack strategy prompt (ASP)** θ in language space, governing how the attack is iterated against a victim VLM, and the **attacker weights** φ, determining the success rate attainable under a given θ.Evolving over a *group of multimodal attack trajectories*, an LLM-based critique of failed trajectories first refines θ, after which group-aggregated attack-success-rate (ASR) rewards update φ. On MM-SafetyBench, MAMJ attains 81.0%, 78.9%, and 82.3% ASR against GPT-4o, Gemini-3-Pro-Preview, and Seed 2.0, exceeding the strongest sample-level baseline by up to 24.1 percentage points, and the learned attacker (θ^\star, φ^\star) transfers *without retraining* to unseen victims while remaining effective under representative defences.These results expose a systemic vulnerability of frontier VLMs to meta-adaptive jailbreaks and underscore the need for defences that anticipate meta-level adversaries. Code and checkpoints will be released.
%U https://aclanthology.org/2026.emnlp-main.258/
%P 5848-5874
Markdown (Informal)
[Fully Unleashing the Multimodal Attacker: Meta-Adaptive Jailbreaking of Vision-Language Models](https://aclanthology.org/2026.emnlp-main.258/) (Pang et al., EMNLP 2026)
ACL
- Shen Pang, Benlei Cui, Yuke Wang, Xuemei Dong, Yuwen Zhai, Jingqun Tang, Haiyang Yu, Hui Xue, Longtao Huang, and Haiwen Hong. 2026. Fully Unleashing the Multimodal Attacker: Meta-Adaptive Jailbreaking of Vision-Language Models. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 5848–5874, Budapest, Hungary. Association for Computational Linguistics.