HarmonyGuard: Toward Safety and Utility in Web Agents via Adaptive Policy Enhancement and Dual-Objective Optimization

Summary

This paper talks about HarmonyGuard, a system that helps web agents created with large language models work better and safer by using multiple agents that focus on following security rules and completing tasks effectively at the same time.

What's the problem?

The problem is that web agents face challenges because the internet is full of hidden threats that change all the time. It's hard to keep these agents safe from risks while making sure they do their jobs well. Current methods usually focus on either safety or task success, but not both, which can lead to problems.

What's the solution?

The solution was to build a multi-agent framework where one agent manages security policies by extracting rules from documents and updating them as threats change, and another agent focuses on balancing safety and task performance using advanced reasoning. These agents work together to constantly improve both protecting users and completing tasks efficiently.

Why it matters?

This matters because it makes web agents more trustworthy and capable, reducing risks while still getting work done. By keeping users safe and improving task success, HarmonyGuard helps AI-driven web tools become more useful and reliable for everyday use.