Image-Based Prompt Injection: Adversarial Images Hijack Multimodal LLMs with 64% Success Rate

arXiv 2603.03637·high signal

Researchers demonstrate Image-based Prompt Injection (IPI), a black-box attack embedding adversarial instructions into natural images to override multimodal LLM behavior without any model access. The most effective configuration achieves up to 64% attack success under stealth constraints across tested VLMs. This establishes a new attack surface distinct from text-based prompt injection, relevant to any pipeline that processes user-supplied images with a VLM.