AI TrackerFrontier desk · AEST
Back to news
normalaitech

AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model

arXiv cs.AI · 7 Oct 2026, 03:56 AEST

AdvSim2Real introduces a new method for training web agents to handle adaptive prompt injections in a web world model. This approach enhances the agent's capabilities and robustness against unseen adversaries.

Key facts

Read aloud

AdvSim2Real introduces a new method for training web agents to handle adaptive prompt injections in a web world model. This approach enhances the agent's capabilities and robustness against unseen adversaries. Web agents complete user requests by reading and acting on pages that third parties write, so an instruction planted on a page can redirect the agent away from the user's goal. Current defenses fine-tune the agent on injections fixed before training, and attackers that adapt to the trained model bypass them. Training in the simulator makes a 4B agent both more capable and more robust: its completion rises with and without attacks, holds against a frontier-model adversary it never trained against, and its capability gain carries over to a real browser. On one hundred fifty web tasks, AdvSim2Real raises completion under this unseen adversary by thirty three point six percent relative to the base agent.

Sources in this cluster