Attributional Safety Failures in Large Language Models under Code-Mixed Perturbations

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Banerjee, Somnath, Chatterjee, Pratyush, Kumar, Shanu, Layek, Sayan, Agrawal, Parag, Hazra, Rima, Mukherjee, Animesh
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918223277981696
author Banerjee, Somnath
Chatterjee, Pratyush
Kumar, Shanu
Layek, Sayan
Agrawal, Parag
Hazra, Rima
Mukherjee, Animesh
author_facet Banerjee, Somnath
Chatterjee, Pratyush
Kumar, Shanu
Layek, Sayan
Agrawal, Parag
Hazra, Rima
Mukherjee, Animesh
contents While LLMs appear robustly safety-aligned in English, we uncover a catastrophic, overlooked weakness: attributional collapse under code-mixed perturbations. Our systematic evaluation of open models shows that the linguistic camouflage of code-mixing -- ``blending languages within a single conversation'' -- can cause safety guardrails to fail dramatically. Attack success rates (ASR) spike from a benign 9\% in monolingual English to 69\% under code-mixed inputs, with rates exceeding 90\% in non-Western contexts such as Arabic and Hindi. These effects hold not only on controlled synthetic datasets but also on real-world social media traces, revealing a serious risk for billions of users. To explain why this happens, we introduce saliency drift attribution (SDA), an interpretability framework that shows how, under code-mixing, the model's internal attention drifts away from safety-critical tokens (e.g., ``violence'' or ``corruption''), effectively blinding it to harmful intent. Finally, we propose a lightweight translation-based restoration strategy that recovers roughly 80\% of the safety lost to code-mixing, offering a practical path toward more equitable and robust LLM safety.
format Preprint
id arxiv_https___arxiv_org_abs_2505_14469
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Attributional Safety Failures in Large Language Models under Code-Mixed Perturbations
Banerjee, Somnath
Chatterjee, Pratyush
Kumar, Shanu
Layek, Sayan
Agrawal, Parag
Hazra, Rima
Mukherjee, Animesh
Computation and Language
Artificial Intelligence
While LLMs appear robustly safety-aligned in English, we uncover a catastrophic, overlooked weakness: attributional collapse under code-mixed perturbations. Our systematic evaluation of open models shows that the linguistic camouflage of code-mixing -- ``blending languages within a single conversation'' -- can cause safety guardrails to fail dramatically. Attack success rates (ASR) spike from a benign 9\% in monolingual English to 69\% under code-mixed inputs, with rates exceeding 90\% in non-Western contexts such as Arabic and Hindi. These effects hold not only on controlled synthetic datasets but also on real-world social media traces, revealing a serious risk for billions of users. To explain why this happens, we introduce saliency drift attribution (SDA), an interpretability framework that shows how, under code-mixing, the model's internal attention drifts away from safety-critical tokens (e.g., ``violence'' or ``corruption''), effectively blinding it to harmful intent. Finally, we propose a lightweight translation-based restoration strategy that recovers roughly 80\% of the safety lost to code-mixing, offering a practical path toward more equitable and robust LLM safety.
title Attributional Safety Failures in Large Language Models under Code-Mixed Perturbations
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2505.14469