Reproduce “AdvSim2Real : Training Web Agents Against Adaptive Prompt Injection in a Web World Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sarim Hashmi, Mukul Ranjan, Kshitij Mishra, Mikhail Kuznetsov, Praneeth Vepakomma, Nils Lukas
AdvSim2Real trains a task curriculum, attacker, and web agent together in a frozen simulator. It rewards attacks only when they turn a judge-accepted success into failure. On 150 benchmark tasks, judged completion against an unseen attacker rises from 23.00% to 30.72%, a reported 33.6% relative gain. These attacked results remain simulator judgments, and task feasibility after injection is unchecked.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.