This research introduces a systematic approach to assess the real-world consequences of AI agents operating mobile interfaces, focusing on identifying potentially risky actions.

• Develops a comprehensive impact taxonomy for evaluating the safety implications of AI agent actions
• Analyzes potential security vulnerabilities when AI agents perform irreversible operations
• Proposes risk assessment methods to prevent harmful consequences before they occur
• Establishes foundations for safer autonomous agents that can navigate mobile UIs responsibly

This work addresses critical security concerns as AI agents become more prevalent in managing daily tasks, helping businesses develop safer AI tools that can operate user interfaces without creating unexpected risks.

From Interaction to Impact: Towards Safer AI Agents Through Understanding and Evaluating Mobile UI Operation Impacts