@inproceedings{pit-etal-2026-one,
title = "One Refusal Isn{'}t Enough: Multi-Turn Jailbreak Attacks on Large Language Models",
author = "Pit, Pagnarasmey and
Ma, Xingjun and
Bailey, James and
Hovy, Eduard",
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.2577/",
pages = "58571--58591",
ISBN = "979-8-89176-459-0",
abstract = "A model that refuses once may not refuse again. We introduce MAZE (Multi-turn Adversarial Zone Exploration), a black-box jailbreak framework that adapts its prompts to the target model{'}s responses across a dialogue. Across ten state-of-the-art models, MAZE achieves 50{\%} attack success on commercial systems and 74{\%} on open-source systems, with the highest observed average ASR and average rank among evaluated baselines. Ablations show that accumulated conversational context is the primary driver: removing target-side history reduces ASR by 32.9{\%} on open-source models and 56.1{\%} on commercial models, revealing strongly path-dependent safety behaviour. When refusals occur, adaptive recovery helps restore conversational progress through reframing. We identify three recurring interaction patterns: legitimacy construction, context-dependent risk calibration, and helpfulness{--}safety tension, that help explain why follow-up requests can receive different safety treatment from isolated requests. Claude-Sonnet-4.5 is a notable exception, maintaining refusals across most reframing attempts. These findings motivate evaluating alignment across full multi-turn trajectories rather than at the point of initial refusal."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="pit-etal-2026-one">
<titleInfo>
<title>One Refusal Isn’t Enough: Multi-Turn Jailbreak Attacks on Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pagnarasmey</namePart>
<namePart type="family">Pit</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xingjun</namePart>
<namePart type="family">Ma</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">James</namePart>
<namePart type="family">Bailey</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Eduard</namePart>
<namePart type="family">Hovy</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>A model that refuses once may not refuse again. We introduce MAZE (Multi-turn Adversarial Zone Exploration), a black-box jailbreak framework that adapts its prompts to the target model’s responses across a dialogue. Across ten state-of-the-art models, MAZE achieves 50% attack success on commercial systems and 74% on open-source systems, with the highest observed average ASR and average rank among evaluated baselines. Ablations show that accumulated conversational context is the primary driver: removing target-side history reduces ASR by 32.9% on open-source models and 56.1% on commercial models, revealing strongly path-dependent safety behaviour. When refusals occur, adaptive recovery helps restore conversational progress through reframing. We identify three recurring interaction patterns: legitimacy construction, context-dependent risk calibration, and helpfulness–safety tension, that help explain why follow-up requests can receive different safety treatment from isolated requests. Claude-Sonnet-4.5 is a notable exception, maintaining refusals across most reframing attempts. These findings motivate evaluating alignment across full multi-turn trajectories rather than at the point of initial refusal.</abstract>
<identifier type="citekey">pit-etal-2026-one</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2577/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>58571</start>
<end>58591</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T One Refusal Isn’t Enough: Multi-Turn Jailbreak Attacks on Large Language Models
%A Pit, Pagnarasmey
%A Ma, Xingjun
%A Bailey, James
%A Hovy, Eduard
%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 pit-etal-2026-one
%X A model that refuses once may not refuse again. We introduce MAZE (Multi-turn Adversarial Zone Exploration), a black-box jailbreak framework that adapts its prompts to the target model’s responses across a dialogue. Across ten state-of-the-art models, MAZE achieves 50% attack success on commercial systems and 74% on open-source systems, with the highest observed average ASR and average rank among evaluated baselines. Ablations show that accumulated conversational context is the primary driver: removing target-side history reduces ASR by 32.9% on open-source models and 56.1% on commercial models, revealing strongly path-dependent safety behaviour. When refusals occur, adaptive recovery helps restore conversational progress through reframing. We identify three recurring interaction patterns: legitimacy construction, context-dependent risk calibration, and helpfulness–safety tension, that help explain why follow-up requests can receive different safety treatment from isolated requests. Claude-Sonnet-4.5 is a notable exception, maintaining refusals across most reframing attempts. These findings motivate evaluating alignment across full multi-turn trajectories rather than at the point of initial refusal.
%U https://aclanthology.org/2026.emnlp-main.2577/
%P 58571-58591
Markdown (Informal)
[One Refusal Isn’t Enough: Multi-Turn Jailbreak Attacks on Large Language Models](https://aclanthology.org/2026.emnlp-main.2577/) (Pit et al., EMNLP 2026)
ACL