Reproduce “Empirical Analysis of Large Vision-Language Models against Goal Hijacking via Visual Prompt Injection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Subaru Kimura, Ryota Tanaka, Shumpei Miyawaki, Jun Suzuki, Keisuke Sakaguchi
The authors evaluate visual prompt injection that places instructions in an image to redirect a vision-language model from the user’s text request to an attacker-chosen task. They test 500 cases across five models, defining success as answering only the target task and answering it correctly. GPT-4V reaches 15.8% success, the highest rate, in this controlled evaluation—not a real-world attack rate.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.