What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks

arXiv·medium signal

LLM-powered content moderation operates on tokenized text and largely ignores visual cues that humans rely on, opening an attack surface where perturbations look benign to the model but carry clear meaning to a human reader. Demonstrates adversarial text attacks that exploit this perception gap. A concrete warning for builders relying on LLM moderation as a primary safety layer.