Announcement_7
New preprint: AdvSim2Real trains web agents against adaptive prompt injection by co-evolving a task curriculum, an injection adversary, and the agent inside a web world model. A 4B agent gains 33.6% relative completion under an unseen frontier-model adversary.