Research
MPIDetector: Detecting Multi-modal Prompt Injection via LLM-based Reasoning
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 25 Oct 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
MPIDetector provides evaluation evidence for detecting cross-modal and web-agent prompt injections through reasoning-based screening without retraining a dedicated classifier.
MPIDetector is a proposed prompt-injection detection framework that uses LLM reasoning, few-shot examples, and prompt templates to assess textual and visual inputs without fine-tuning. The paper reports up to 90.91% accuracy on textual tasks, 99.28% accuracy on visual tasks, and a 99.80% detection rate against 500 obfuscated WebInject web-agent injections.