Misaligned Roles, Misplaced Images: Structural Input Perturbations Expose Multimodal Alignment Blind Spots

Erfan Shayegani, G M Shahariar, Sara Abdali, Lei Yu, Nael Abu-Ghazaleh, Yue Dong, [5pt] University of California, Riverside, Microsoft Applied Sciences Group, University of Toronto[5pt] {sshay004, gshah010, naelag, yued}@ucr.edu · 2025

Read Misaligned Roles, Misplaced Images: Structural Input Perturbations Expose Multimodal Alignment Blind Spots by Erfan Shayegani, G M Shahariar, Sara Abdali, Lei Yu, Nael Abu-Ghazaleh, Yue Dong, [5pt] University of California, Riverside, Microsoft Applied Sciences Group, University of Toronto[5pt] {sshay004, gshah010, naelag, yued}@ucr.edu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-03735.txt.

Description

Erfan Shayegani and co-authors show that shuffling the structure of a prompt, not its words, can break safety training in image-language models, in work from 2025. Safety tuning mostly polices the assistant's seat. The user's seat and the fixed token layout stay untested.

Their attacks swap roles so text lands in the wrong seat, or move the image token out of place. The query itself stays untouched. Across eight settings and several models, the tricks drew out harmful output, and stacking them hit harder. Internal traces leaned toward non-refusal, as in known jailbreaks. A training fix that mixes scrambled harmful and benign prompts made models sturdier. The work flags structure as a blind spot.

Catalog record

CREATORErfan Shayegani, G M Shahariar, Sara Abdali, Lei Yu, Nael Abu-Ghazaleh, Yue Dong, [5pt] University of California, Riverside, Microsoft Applied Sciences Group, University of Toronto[5pt] {sshay004, gshah010, naelag, yued}@ucr.edu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE72 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-03735

Browse the full library · About the Tower of Babel