Image-Based Prompt Injection: Typographic Adversarial Instructions on Physical Objects Hijack Multimodal Agents Without Any Digital File Manipulation

Cloud Security Alliance·high signal

CSA Research documents a novel threat class where adversarial text embedded on physical objects (stickers, printed signs, whiteboards) in photos can hijack multimodal LLM agent behavior. Unlike digital injection, there is no file to inspect or sanitize — the attack vector is the physical world captured by camera. This is particularly relevant for agents with computer-use or vision capabilities operating in real-world environments.