AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model
Presents novel adversarial training methodology for web agent security with significant performance improvements.
AI Summary
AdvSim2Real trains web agents against adaptive injection attacks using co-evolution of tasks, adversaries, and agents in a web world model.
Excerpt
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. The agent cannot simply ignore the page, because the page also holds the values and controls the task requires. Current defenses fine-tune the agent on injections fixed before training, and attackers that adapt to the trained model bypass them. Adversarial training lets the attacker adapt but keeps the tasks fixed, so a t