A new paper proposes a shift from the prevalent ReAct paradigm to a compile-and-execute approach for web agents. The authors argue that ReAct's direct flow of runtime web content into the model increases vulnerability to prompt injections. A compile-and-execute model, conversely, commits to a task-specific program before observing web content, thereby isolating untrusted data within a predefined execution graph. The paper analyzes the WebArena benchmark, finding that most tasks are compatible with this proposed paradigm, with a significant portion executable without runtime LLM subroutines. The primary obstacle identified is the lack of semantic, typed interfaces for web interactions, which currently rely on generic browser actions like clicks and scrolls. AI
IMPACT This research could enhance the security and reliability of web-based AI agents by mitigating prompt injection risks.
RANK_REASON Academic paper proposing a new paradigm for AI agents. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
