@inproceedings{qin-feng-2026-proactive,
title = "Proactive Safety Alignment: Enabling Proactive Initiative in Multi-Turn Defence",
author = "Qin, Mengyuan and
Feng, Yue",
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.2208/",
pages = "47300--47311",
ISBN = "979-8-89176-462-0",
abstract = "LLMs are increasingly deployed as interactive agents capable of multi-turn dialogues and tool usage, but they remain vulnerable to adaptive attacks. Existing safety mechanisms are mostly reactive, only attempting to uncover pre-existing hidden malicious intent, which makes them largely ineffective against attacks when such attacks are inherently designed to conceal their intent, and causes them to often fail in multi-turn, interactive, or out-of-distribution scenarios. In addition, existing multi-turn defences are largely data-driven, which limits their generalisation ability. To address this challenge, we propose Proactive Safety Alignment (PSA), a framework that anticipates user intent proactively to align LLM agents' outputs with safety objectives. PSA combines Safe User Request Prediction (SURP), which anticipates likely safe future requests, and Potentially Harmful Generation (PHG), which evaluates whether the current interaction state can support misuse. Evaluations across multiple benchmarks demonstrate that PSA substantially reduces attack success rates, while preserving overall utility on MT-Bench. Our results show that PSA provides an effective and scalable paradigm for multi-turn LLM safety alignment. The code is available at https://github.com/qinmengyuan/PSA."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="qin-feng-2026-proactive">
<titleInfo>
<title>Proactive Safety Alignment: Enabling Proactive Initiative in Multi-Turn Defence</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mengyuan</namePart>
<namePart type="family">Qin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yue</namePart>
<namePart type="family">Feng</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>LLMs are increasingly deployed as interactive agents capable of multi-turn dialogues and tool usage, but they remain vulnerable to adaptive attacks. Existing safety mechanisms are mostly reactive, only attempting to uncover pre-existing hidden malicious intent, which makes them largely ineffective against attacks when such attacks are inherently designed to conceal their intent, and causes them to often fail in multi-turn, interactive, or out-of-distribution scenarios. In addition, existing multi-turn defences are largely data-driven, which limits their generalisation ability. To address this challenge, we propose Proactive Safety Alignment (PSA), a framework that anticipates user intent proactively to align LLM agents’ outputs with safety objectives. PSA combines Safe User Request Prediction (SURP), which anticipates likely safe future requests, and Potentially Harmful Generation (PHG), which evaluates whether the current interaction state can support misuse. Evaluations across multiple benchmarks demonstrate that PSA substantially reduces attack success rates, while preserving overall utility on MT-Bench. Our results show that PSA provides an effective and scalable paradigm for multi-turn LLM safety alignment. The code is available at https://github.com/qinmengyuan/PSA.</abstract>
<identifier type="citekey">qin-feng-2026-proactive</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2208/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>47300</start>
<end>47311</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Proactive Safety Alignment: Enabling Proactive Initiative in Multi-Turn Defence
%A Qin, Mengyuan
%A Feng, Yue
%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 qin-feng-2026-proactive
%X LLMs are increasingly deployed as interactive agents capable of multi-turn dialogues and tool usage, but they remain vulnerable to adaptive attacks. Existing safety mechanisms are mostly reactive, only attempting to uncover pre-existing hidden malicious intent, which makes them largely ineffective against attacks when such attacks are inherently designed to conceal their intent, and causes them to often fail in multi-turn, interactive, or out-of-distribution scenarios. In addition, existing multi-turn defences are largely data-driven, which limits their generalisation ability. To address this challenge, we propose Proactive Safety Alignment (PSA), a framework that anticipates user intent proactively to align LLM agents’ outputs with safety objectives. PSA combines Safe User Request Prediction (SURP), which anticipates likely safe future requests, and Potentially Harmful Generation (PHG), which evaluates whether the current interaction state can support misuse. Evaluations across multiple benchmarks demonstrate that PSA substantially reduces attack success rates, while preserving overall utility on MT-Bench. Our results show that PSA provides an effective and scalable paradigm for multi-turn LLM safety alignment. The code is available at https://github.com/qinmengyuan/PSA.
%U https://aclanthology.org/2026.findings-emnlp.2208/
%P 47300-47311
Markdown (Informal)
[Proactive Safety Alignment: Enabling Proactive Initiative in Multi-Turn Defence](https://aclanthology.org/2026.findings-emnlp.2208/) (Qin & Feng, Findings 2026)
ACL