@inproceedings{rao-etal-2026-mmjail,
title = "{MMJ}ail: Compounded Multilingual and Multi-Turn Vulnerabilities in {LLM} Bias Resistance",
author = "Rao, Anisha and
Kim, Takyoung and
Hakkani-Tur, Dilek",
editor = "Calabrese, Agostina and
Davidson, Thomas and
de Kock, Christine and
Khurana, Urja and
Marchiori Manerba, Marta and
Piot, Paloma and
Talat, Zeerak",
booktitle = "Proceedings of the 10th Workshop on Online Abuse and Harms ({WOAH})",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.woah-1.36/",
pages = "434--449",
ISBN = "979-8-89176-487-3",
abstract = "LLM safety evaluation remains largely English-centric and single-turn, despite real-world use in multilingual, multi-turn dialogue. We introduce MMJail, a controlled benchmark that embeds bias-eliciting prompts into controlled five-turn escalations while keeping the final prompt identical to the single-turn baseline. This design isolates the effect of conversational context across six languages: English, Vietnamese, Arabic, Thai, Bengali, and Swahili. Across Llama-3.1-8B, Qwen3.5-9B, and Gemma-3-12B, multi-turn escalation disproportionately increases attack success in lower-resource languages; for example, Qwen3.5-9B{'}s Bengali ASR rises from 17.6{\%} to 66.7{\%}. Turn-level analysis shows that escalation creates new jailbreaks and reveals distinct failure modes, including gradual drift, final-turn capitulation, and mid-conversation recovery. Along with behavioral analyses from multiple perspectives, we conduct a preliminary fine-tuning mitigation to reduce jailbreaking in several languages, suggesting that multilingual multi-turn safety can be improved through targeted supervision."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="rao-etal-2026-mmjail">
<titleInfo>
<title>MMJail: Compounded Multilingual and Multi-Turn Vulnerabilities in LLM Bias Resistance</title>
</titleInfo>
<name type="personal">
<namePart type="given">Anisha</namePart>
<namePart type="family">Rao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Takyoung</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dilek</namePart>
<namePart type="family">Hakkani-Tur</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 10th Workshop on Online Abuse and Harms (WOAH)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Agostina</namePart>
<namePart type="family">Calabrese</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Thomas</namePart>
<namePart type="family">Davidson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christine</namePart>
<namePart type="family">de Kock</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Urja</namePart>
<namePart type="family">Khurana</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marta</namePart>
<namePart type="family">Marchiori Manerba</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Paloma</namePart>
<namePart type="family">Piot</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zeerak</namePart>
<namePart type="family">Talat</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">979-8-89176-487-3</identifier>
</relatedItem>
<abstract>LLM safety evaluation remains largely English-centric and single-turn, despite real-world use in multilingual, multi-turn dialogue. We introduce MMJail, a controlled benchmark that embeds bias-eliciting prompts into controlled five-turn escalations while keeping the final prompt identical to the single-turn baseline. This design isolates the effect of conversational context across six languages: English, Vietnamese, Arabic, Thai, Bengali, and Swahili. Across Llama-3.1-8B, Qwen3.5-9B, and Gemma-3-12B, multi-turn escalation disproportionately increases attack success in lower-resource languages; for example, Qwen3.5-9B’s Bengali ASR rises from 17.6% to 66.7%. Turn-level analysis shows that escalation creates new jailbreaks and reveals distinct failure modes, including gradual drift, final-turn capitulation, and mid-conversation recovery. Along with behavioral analyses from multiple perspectives, we conduct a preliminary fine-tuning mitigation to reduce jailbreaking in several languages, suggesting that multilingual multi-turn safety can be improved through targeted supervision.</abstract>
<identifier type="citekey">rao-etal-2026-mmjail</identifier>
<location>
<url>https://aclanthology.org/2026.woah-1.36/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>434</start>
<end>449</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T MMJail: Compounded Multilingual and Multi-Turn Vulnerabilities in LLM Bias Resistance
%A Rao, Anisha
%A Kim, Takyoung
%A Hakkani-Tur, Dilek
%Y Calabrese, Agostina
%Y Davidson, Thomas
%Y de Kock, Christine
%Y Khurana, Urja
%Y Marchiori Manerba, Marta
%Y Piot, Paloma
%Y Talat, Zeerak
%S Proceedings of the 10th Workshop on Online Abuse and Harms (WOAH)
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-487-3
%F rao-etal-2026-mmjail
%X LLM safety evaluation remains largely English-centric and single-turn, despite real-world use in multilingual, multi-turn dialogue. We introduce MMJail, a controlled benchmark that embeds bias-eliciting prompts into controlled five-turn escalations while keeping the final prompt identical to the single-turn baseline. This design isolates the effect of conversational context across six languages: English, Vietnamese, Arabic, Thai, Bengali, and Swahili. Across Llama-3.1-8B, Qwen3.5-9B, and Gemma-3-12B, multi-turn escalation disproportionately increases attack success in lower-resource languages; for example, Qwen3.5-9B’s Bengali ASR rises from 17.6% to 66.7%. Turn-level analysis shows that escalation creates new jailbreaks and reveals distinct failure modes, including gradual drift, final-turn capitulation, and mid-conversation recovery. Along with behavioral analyses from multiple perspectives, we conduct a preliminary fine-tuning mitigation to reduce jailbreaking in several languages, suggesting that multilingual multi-turn safety can be improved through targeted supervision.
%U https://aclanthology.org/2026.woah-1.36/
%P 434-449
Markdown (Informal)
[MMJail: Compounded Multilingual and Multi-Turn Vulnerabilities in LLM Bias Resistance](https://aclanthology.org/2026.woah-1.36/) (Rao et al., WOAH 2026)
ACL