@inproceedings{aidakhmetov-etal-2026-steering,
title = "Steering Vectors are an Adversarial Attack Surface",
author = "Aidakhmetov, Abzal and
Crisostomi, Donato and
Mencattini, Tommaso and
Minut, Adrian Robert and
Masi, Iacopo and
Rodol{\`a}, Emanuele",
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.2040/",
pages = "43420--43439",
ISBN = "979-8-89176-462-0",
abstract = "Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a stealth data poisoning attack silently compromises this pipeline. By substituting $4{-}6\%$ of tokens in the steering dataset, an attacker can silently align the resulting vector with an anti-refusal direction. This jailbreaks the target model while preserving the intended steering effect on benign prompts. Under this threat model, a malicious actor can distribute an apparently safe bundle containing texts, vectors, and weights, alongside an equivalence certificate that the end-user can verify. We test the attack on two open-weight model families and eight model-attribute combinations, observing that poisoned vectors reach an absolute attack success rate (ASR) of $20{-}55\%$, $+19\%$ to $+51\%$ over a clean reference. The vulnerability is not an artefact of scale or of access to the victim: it persists natively at 70B, and a poisoning optimised once on a small surrogate transfers to unseen models that recompute their own vectors from the shared texts. Finally, we find that a refusal-direction orthogonalization defense can recover ${\approx}82\%$ of the ASR gap without harming benign behavior."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="aidakhmetov-etal-2026-steering">
<titleInfo>
<title>Steering Vectors are an Adversarial Attack Surface</title>
</titleInfo>
<name type="personal">
<namePart type="given">Abzal</namePart>
<namePart type="family">Aidakhmetov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Donato</namePart>
<namePart type="family">Crisostomi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tommaso</namePart>
<namePart type="family">Mencattini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Adrian</namePart>
<namePart type="given">Robert</namePart>
<namePart type="family">Minut</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Iacopo</namePart>
<namePart type="family">Masi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Emanuele</namePart>
<namePart type="family">Rodolà</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>Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a stealth data poisoning attack silently compromises this pipeline. By substituting 4-6% of tokens in the steering dataset, an attacker can silently align the resulting vector with an anti-refusal direction. This jailbreaks the target model while preserving the intended steering effect on benign prompts. Under this threat model, a malicious actor can distribute an apparently safe bundle containing texts, vectors, and weights, alongside an equivalence certificate that the end-user can verify. We test the attack on two open-weight model families and eight model-attribute combinations, observing that poisoned vectors reach an absolute attack success rate (ASR) of 20-55%, +19% to +51% over a clean reference. The vulnerability is not an artefact of scale or of access to the victim: it persists natively at 70B, and a poisoning optimised once on a small surrogate transfers to unseen models that recompute their own vectors from the shared texts. Finally, we find that a refusal-direction orthogonalization defense can recover \approx82% of the ASR gap without harming benign behavior.</abstract>
<identifier type="citekey">aidakhmetov-etal-2026-steering</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.2040/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>43420</start>
<end>43439</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Steering Vectors are an Adversarial Attack Surface
%A Aidakhmetov, Abzal
%A Crisostomi, Donato
%A Mencattini, Tommaso
%A Minut, Adrian Robert
%A Masi, Iacopo
%A Rodolà, Emanuele
%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 aidakhmetov-etal-2026-steering
%X Activation steering has become a popular way to control Large Language Model (LLM) behavior without fine-tuning. Since the technique is plug-and-play, users share datasets and precomputed vectors to steer model activations. However, we show that a stealth data poisoning attack silently compromises this pipeline. By substituting 4-6% of tokens in the steering dataset, an attacker can silently align the resulting vector with an anti-refusal direction. This jailbreaks the target model while preserving the intended steering effect on benign prompts. Under this threat model, a malicious actor can distribute an apparently safe bundle containing texts, vectors, and weights, alongside an equivalence certificate that the end-user can verify. We test the attack on two open-weight model families and eight model-attribute combinations, observing that poisoned vectors reach an absolute attack success rate (ASR) of 20-55%, +19% to +51% over a clean reference. The vulnerability is not an artefact of scale or of access to the victim: it persists natively at 70B, and a poisoning optimised once on a small surrogate transfers to unseen models that recompute their own vectors from the shared texts. Finally, we find that a refusal-direction orthogonalization defense can recover \approx82% of the ASR gap without harming benign behavior.
%U https://aclanthology.org/2026.findings-emnlp.2040/
%P 43420-43439
Markdown (Informal)
[Steering Vectors are an Adversarial Attack Surface](https://aclanthology.org/2026.findings-emnlp.2040/) (Aidakhmetov et al., Findings 2026)
ACL
- Abzal Aidakhmetov, Donato Crisostomi, Tommaso Mencattini, Adrian Robert Minut, Iacopo Masi, and Emanuele Rodolà. 2026. Steering Vectors are an Adversarial Attack Surface. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 43420–43439, Budapest, Hungary. Association for Computational Linguistics.