@inproceedings{im-etal-2026-legible,
title = "Legible but Not Localized: Jailbreak Geometry in Mixture-of-Experts Language Models Is Attack-Specific",
author = "Im, Yubin and
Kim, Myeonghyun and
Kim, Seongmin and
Cho, Haehyun",
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.2049/",
pages = "46673--46687",
ISBN = "979-8-89176-459-0",
abstract = "Activation-based jailbreak defenses rest on a fixed assumption: a successful jailbreak leaves an internal trace stable enough across attacks for one probe to catch new ones. We show it fails in mixture-of-experts language models. Within an attack, routing makes the jailbreak legible: router logits distinguish compliance from refusal, and a few experts carry the signal. But the signature does not survive across attacks. A routing probe trained on three attacks performs at chance on a held-out fourth (pooled leave-one-attack-out balanced accuracy 0.52, 95 percent CI [0.50, 0.61]). Controls on two dense models show near-chance residual-stream transfer to held-out roleplay, so the failure is not specific to MoE routing. Nor is it an artifact of separable regions of activation space: the comply-versus-refuse direction itself shifts across attacks at every network depth. One systematic trace remains in the compliance state{'}s effective dimensionality, but it is a signature of shape, not location, beyond directional probes. Across six open-weight MoE models and four jailbreak attacks, in routing and residual streams alike, the jailbroken state is legible and locally structured within each attack, yet its location is attack-specific. Reading activations alone cannot be assumed to catch a jailbreak the probe was not built for."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="im-etal-2026-legible">
<titleInfo>
<title>Legible but Not Localized: Jailbreak Geometry in Mixture-of-Experts Language Models Is Attack-Specific</title>
</titleInfo>
<name type="personal">
<namePart type="given">Yubin</namePart>
<namePart type="family">Im</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Myeonghyun</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seongmin</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Haehyun</namePart>
<namePart type="family">Cho</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>Activation-based jailbreak defenses rest on a fixed assumption: a successful jailbreak leaves an internal trace stable enough across attacks for one probe to catch new ones. We show it fails in mixture-of-experts language models. Within an attack, routing makes the jailbreak legible: router logits distinguish compliance from refusal, and a few experts carry the signal. But the signature does not survive across attacks. A routing probe trained on three attacks performs at chance on a held-out fourth (pooled leave-one-attack-out balanced accuracy 0.52, 95 percent CI [0.50, 0.61]). Controls on two dense models show near-chance residual-stream transfer to held-out roleplay, so the failure is not specific to MoE routing. Nor is it an artifact of separable regions of activation space: the comply-versus-refuse direction itself shifts across attacks at every network depth. One systematic trace remains in the compliance state’s effective dimensionality, but it is a signature of shape, not location, beyond directional probes. Across six open-weight MoE models and four jailbreak attacks, in routing and residual streams alike, the jailbroken state is legible and locally structured within each attack, yet its location is attack-specific. Reading activations alone cannot be assumed to catch a jailbreak the probe was not built for.</abstract>
<identifier type="citekey">im-etal-2026-legible</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2049/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>46673</start>
<end>46687</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Legible but Not Localized: Jailbreak Geometry in Mixture-of-Experts Language Models Is Attack-Specific
%A Im, Yubin
%A Kim, Myeonghyun
%A Kim, Seongmin
%A Cho, Haehyun
%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 im-etal-2026-legible
%X Activation-based jailbreak defenses rest on a fixed assumption: a successful jailbreak leaves an internal trace stable enough across attacks for one probe to catch new ones. We show it fails in mixture-of-experts language models. Within an attack, routing makes the jailbreak legible: router logits distinguish compliance from refusal, and a few experts carry the signal. But the signature does not survive across attacks. A routing probe trained on three attacks performs at chance on a held-out fourth (pooled leave-one-attack-out balanced accuracy 0.52, 95 percent CI [0.50, 0.61]). Controls on two dense models show near-chance residual-stream transfer to held-out roleplay, so the failure is not specific to MoE routing. Nor is it an artifact of separable regions of activation space: the comply-versus-refuse direction itself shifts across attacks at every network depth. One systematic trace remains in the compliance state’s effective dimensionality, but it is a signature of shape, not location, beyond directional probes. Across six open-weight MoE models and four jailbreak attacks, in routing and residual streams alike, the jailbroken state is legible and locally structured within each attack, yet its location is attack-specific. Reading activations alone cannot be assumed to catch a jailbreak the probe was not built for.
%U https://aclanthology.org/2026.emnlp-main.2049/
%P 46673-46687
Markdown (Informal)
[Legible but Not Localized: Jailbreak Geometry in Mixture-of-Experts Language Models Is Attack-Specific](https://aclanthology.org/2026.emnlp-main.2049/) (Im et al., EMNLP 2026)
ACL