"MirageAttack"

" MirageWeb is a benchmark designed to evaluate how screenshot based AI agents behave when critical on screen information is hidden by Mirage Attack. It includes 20 realistic web scenarios covering both information retrieval and interactive tasks, such as reading financial data, checking contract clauses, filling forms, confirming payments, and deleting repositories. For each scenario, MirageWeb provides a clean website and a Mirage attacked counterpart, allowing direct comparison between normal and compromised conditions. The benchmark preserves the original page structure and functionality, so hidden elements remain usable for humans on HDR displays while becoming invisible in screenshots captured by agents. This makes MirageWeb useful for studying not only task success but also agent failure modes, reasoning errors, and unsafe actions. In the paper, MirageWeb is used to measure task success rate and completion time across multiple commercial agents, showing that Mirage Attack can substantially disrupt perception and downstream decision making."