Beyond Accuracy: Identifying the Stealth Gap in Multimodal Prompt Injection Defenses via a Four‑Dimensional Taxonomy

Abstract: Prompt injection is the most significant security concern for large language models. When a model reads documents, photos, and retrieved content in addition to text, management becomes more challenging. Instructions can be hidden by an attacker in a tool output, a retrieved source, a document layer, or a visual overlay. Although prompt injection research has expanded rapidly, current studies still lack a standard framework for evaluating attacks and defences which often employ inconsistent threat models and evaluation methods. This review addresses the gap, by classifying attacks according to four dimensions: carrier type, pipeline position, attacker objectives, and stealth level. We analyze defence capabilities throughout the threat landscape using a coverage score and find significant gaps, particularly in protecting against stealth-oriented attacks, which remain understudied.

Keywords: Prompt Injection, LLM, AI Security, Attack Taxonomy, Dimensional Coverage.