Research
A Cross-Modal Prompt Injection Attack against Large Vision-Language Models with Image-Only Perturbation
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 10 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
CrossMPI demonstrates that attacker-controlled images can influence textual as well as visual interpretation, making image inputs a cross-modal attack surface for defenders of multimodal AI systems.
CrossMPI is a cross-modal prompt-injection method that uses image-only perturbations to steer large vision-language models’ interpretation of both textual and visual inputs. The researchers report improved performance over baseline attacks across multiple models and datasets by targeting multimodal hidden states, selecting middle model layers, and concentrating perturbation budgets near semantically important image regions.