@inproceedings{faustini-etal-2026-first,
title = "Before the First Token: Jailbreak Detection with Attention Probes",
author = "Faustini, Pedro and
McIver, Annabelle and
Dras, Mark",
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.2242/",
pages = "48011--48029",
ISBN = "979-8-89176-462-0",
abstract = "Large language models (LLMs) are deployed with safety guardrails, yet jailbreak attacks and malicious prompts keep bypassing them. Re-aligning an LLM each time a new attack appears is more expensive than crafting attacks, so jailbreak detection is increasingly framed as a pre-processing step: a cheap detector filters out a malicious prompt \textit{before} it reaches the LLM. This need is reinforced by multilingual attacks, where simply translating a prompt to another language can bypass the guardrails. Our goal is to find whether internal representations of a frozen LLM already separate malicious from benign prompts, and whether such a signal is resilient against translation. We study two types of detectors that require no fine-tuning of the LLM: a logistic-regression classifier over whole-sentence embeddings, and a probe-based detector that trains one logistic-regression probe per attention head on the attention activations of the last input token. They need at most a single forward pass. Our experiments show that both match or outperform strong sentence-embedding detectors and off-the-shelf baselines, specially under translation. The probes degrade more gracefully on higher-resource languages and remain competitive on lower-resource ones. Our results indicate that a frozen LLM linearly encodes whether an incoming prompt is malicious in a small set of mid-to-late layers. Per-head attention probes thus offer a defence that can be cheaply retrained as new attacks emerge, without touching the underlying model."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="faustini-etal-2026-first">
<titleInfo>
<title>Before the First Token: Jailbreak Detection with Attention Probes</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pedro</namePart>
<namePart type="family">Faustini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Annabelle</namePart>
<namePart type="family">McIver</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mark</namePart>
<namePart type="family">Dras</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>Large language models (LLMs) are deployed with safety guardrails, yet jailbreak attacks and malicious prompts keep bypassing them. Re-aligning an LLM each time a new attack appears is more expensive than crafting attacks, so jailbreak detection is increasingly framed as a pre-processing step: a cheap detector filters out a malicious prompt before it reaches the LLM. This need is reinforced by multilingual attacks, where simply translating a prompt to another language can bypass the guardrails. Our goal is to find whether internal representations of a frozen LLM already separate malicious from benign prompts, and whether such a signal is resilient against translation. We study two types of detectors that require no fine-tuning of the LLM: a logistic-regression classifier over whole-sentence embeddings, and a probe-based detector that trains one logistic-regression probe per attention head on the attention activations of the last input token. They need at most a single forward pass. Our experiments show that both match or outperform strong sentence-embedding detectors and off-the-shelf baselines, specially under translation. The probes degrade more gracefully on higher-resource languages and remain competitive on lower-resource ones. Our results indicate that a frozen LLM linearly encodes whether an incoming prompt is malicious in a small set of mid-to-late layers. Per-head attention probes thus offer a defence that can be cheaply retrained as new attacks emerge, without touching the underlying model.</abstract>
<identifier type="citekey">faustini-etal-2026-first</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2242/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>48011</start>
<end>48029</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Before the First Token: Jailbreak Detection with Attention Probes
%A Faustini, Pedro
%A McIver, Annabelle
%A Dras, Mark
%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 faustini-etal-2026-first
%X Large language models (LLMs) are deployed with safety guardrails, yet jailbreak attacks and malicious prompts keep bypassing them. Re-aligning an LLM each time a new attack appears is more expensive than crafting attacks, so jailbreak detection is increasingly framed as a pre-processing step: a cheap detector filters out a malicious prompt before it reaches the LLM. This need is reinforced by multilingual attacks, where simply translating a prompt to another language can bypass the guardrails. Our goal is to find whether internal representations of a frozen LLM already separate malicious from benign prompts, and whether such a signal is resilient against translation. We study two types of detectors that require no fine-tuning of the LLM: a logistic-regression classifier over whole-sentence embeddings, and a probe-based detector that trains one logistic-regression probe per attention head on the attention activations of the last input token. They need at most a single forward pass. Our experiments show that both match or outperform strong sentence-embedding detectors and off-the-shelf baselines, specially under translation. The probes degrade more gracefully on higher-resource languages and remain competitive on lower-resource ones. Our results indicate that a frozen LLM linearly encodes whether an incoming prompt is malicious in a small set of mid-to-late layers. Per-head attention probes thus offer a defence that can be cheaply retrained as new attacks emerge, without touching the underlying model.
%U https://aclanthology.org/2026.findings-emnlp.2242/
%P 48011-48029
Markdown (Informal)
[Before the First Token: Jailbreak Detection with Attention Probes](https://aclanthology.org/2026.findings-emnlp.2242/) (Faustini et al., Findings 2026)
ACL