Benchmarking Web Agent Safety under E-commerce Deceptive Interfaces
Researchers have developed a new framework called WebDecept to evaluate the safety of autonomous web agents when interacting with deceptive e-commerce interfaces. The study found that current web agents are highly vulnerable to various deceptive patterns, such as targeted ads and domain redirection, and that prompt-based safety constraints are often inadequate. The findings underscore the need for improved safety measures as web agents become more prevalent in real-world applications. AI
IMPACT Highlights critical safety vulnerabilities in current web agents, necessitating improved defenses for real-world deployment.