Reproduce “MPIDetector: Detecting Multi-modal Prompt Injection via LLM-based Reasoning” 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

Yiheng Li, Jingyue Li

MPIDetector asks whether a multimodal language model can detect hidden instructions in webpage screenshots. It uses few-shot prompts that ask models to inspect visual artifacts, extract readable meaning, and compare the two. On 500 WebInject attacks, Gemini-2.0-Flash flagged 499 and falsely flagged none of 500 benign images. This benchmark result does not establish protection in live agent use.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.