Research
AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 5 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
AdvSim2Real addresses the limitations of training web agents against fixed prompt injections by adapting both attacks and task difficulty while measuring robustness and legitimate task performance.
AdvSim2Real is a training method that co-evolves web tasks, a prompt-injection adversary, and a web agent inside a frozen web world model. Across 150 web tasks, the researchers report a 33.6% relative improvement in completion under an unseen adversary compared with the base 4B agent, with capability gains also transferring to a real browser.