@inproceedings{raihan-etal-2026-robustness,
title = "On the Robustness of Code {LLM} Guardrails to Code-mixed and Transliterated Inputs",
author = "Raihan, Nishat and
Meher, Dipak and
Dhingra, Bhuwan and
Zampieri, Marcos",
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.568/",
pages = "11962--11981",
ISBN = "979-8-89176-462-0",
abstract = "Safety guardrails in Code Large Language Models (Code LLMs) are designed to prevent the generation of malicious code, yet their robustness under linguistically diverse inputs remains largely unexplored. To address this important safety issue, we present MALICE (Malicious Adversarial LLM Inputs for Code Generation), a large-scale adversarial benchmark that exposes a critical blind spot at the intersection of multilingual malicious prompts and code generation safety. Our approach converts malicious English prompts into code-mixed and transliterated variants through sentence-level translation into low-resource languages while preserving programming keywords intact. The resulting dataset comprises approximately 250,000 adversarial prompts spanning 18 languages at various resource levels, derived from nine existing malicious code benchmarks. We perform a comprehensive evaluation across nine models (including Claude 4.6 and GPT-5.2) and safety tools, which reveals concerning vulnerabilities: code-mixed malicious prompts reduce detection accuracy by over 0.70 absolute points for the strongest classifier, while even the most safety-aligned models see their malicious code generation rates increase threefold. These results demonstrate that current safety alignment fails to generalize beyond English inputs, underscoring an urgent need for multilingual defenses in Code LLMs."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="raihan-etal-2026-robustness">
<titleInfo>
<title>On the Robustness of Code LLM Guardrails to Code-mixed and Transliterated Inputs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nishat</namePart>
<namePart type="family">Raihan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dipak</namePart>
<namePart type="family">Meher</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bhuwan</namePart>
<namePart type="family">Dhingra</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marcos</namePart>
<namePart type="family">Zampieri</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>Safety guardrails in Code Large Language Models (Code LLMs) are designed to prevent the generation of malicious code, yet their robustness under linguistically diverse inputs remains largely unexplored. To address this important safety issue, we present MALICE (Malicious Adversarial LLM Inputs for Code Generation), a large-scale adversarial benchmark that exposes a critical blind spot at the intersection of multilingual malicious prompts and code generation safety. Our approach converts malicious English prompts into code-mixed and transliterated variants through sentence-level translation into low-resource languages while preserving programming keywords intact. The resulting dataset comprises approximately 250,000 adversarial prompts spanning 18 languages at various resource levels, derived from nine existing malicious code benchmarks. We perform a comprehensive evaluation across nine models (including Claude 4.6 and GPT-5.2) and safety tools, which reveals concerning vulnerabilities: code-mixed malicious prompts reduce detection accuracy by over 0.70 absolute points for the strongest classifier, while even the most safety-aligned models see their malicious code generation rates increase threefold. These results demonstrate that current safety alignment fails to generalize beyond English inputs, underscoring an urgent need for multilingual defenses in Code LLMs.</abstract>
<identifier type="citekey">raihan-etal-2026-robustness</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.568/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>11962</start>
<end>11981</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T On the Robustness of Code LLM Guardrails to Code-mixed and Transliterated Inputs
%A Raihan, Nishat
%A Meher, Dipak
%A Dhingra, Bhuwan
%A Zampieri, Marcos
%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 raihan-etal-2026-robustness
%X Safety guardrails in Code Large Language Models (Code LLMs) are designed to prevent the generation of malicious code, yet their robustness under linguistically diverse inputs remains largely unexplored. To address this important safety issue, we present MALICE (Malicious Adversarial LLM Inputs for Code Generation), a large-scale adversarial benchmark that exposes a critical blind spot at the intersection of multilingual malicious prompts and code generation safety. Our approach converts malicious English prompts into code-mixed and transliterated variants through sentence-level translation into low-resource languages while preserving programming keywords intact. The resulting dataset comprises approximately 250,000 adversarial prompts spanning 18 languages at various resource levels, derived from nine existing malicious code benchmarks. We perform a comprehensive evaluation across nine models (including Claude 4.6 and GPT-5.2) and safety tools, which reveals concerning vulnerabilities: code-mixed malicious prompts reduce detection accuracy by over 0.70 absolute points for the strongest classifier, while even the most safety-aligned models see their malicious code generation rates increase threefold. These results demonstrate that current safety alignment fails to generalize beyond English inputs, underscoring an urgent need for multilingual defenses in Code LLMs.
%U https://aclanthology.org/2026.findings-emnlp.568/
%P 11962-11981
Markdown (Informal)
[On the Robustness of Code LLM Guardrails to Code-mixed and Transliterated Inputs](https://aclanthology.org/2026.findings-emnlp.568/) (Raihan et al., Findings 2026)
ACL