Web agents are AI programs that browse the web, shop, and manage accounts for their users. Just like humans, they can stumble upon fake online stores or investment scams designed to steal money or personal data. This is a growing problem because such fraud costs victims hundreds of billions each year.
The authors built a safe testbed called ScamWorld that recreates hundreds of real scam websites, preserving how they trick people. They tested six advanced language models (LLMs) acting as web agents, letting each one carry out normal tasks on these fake sites to see how often they would be fooled.
The main finding: across models, agents were deceived 11% to 35% of the time. They were more willing to make payments than to give away passwords or wallet details. Even when agents noticed something was wrong, they often proceeded because the user had asked them to complete the task. This shows that current web agents need better safeguards against online fraud.