Large vision-language models (VLMs) often exhibit weakened safety alignment when visual input is added, according to the abstract of a new arXiv paper. Even when a text prompt already contains explicit harmful intent, attaching an image can substantially increase the risk of a jailbreak, the authors report.

The paper's title points to "jailbreak-related representation shift" as the central mechanism. The work aims both to understand how this shift enables attacks and to defend against it, though the abstract provided is truncated and does not yet describe the specific defense approach.

This is a replacement announcement (v2), indicating the paper has been revised. The full details of the proposed defense are not available in the source text.