From Configuration Contact: Referring to Safer Ai's agents with understanding and evaluation the effects of the UI's UI functionality

With the development produced in AI, there is an increase made in creating independent agents that can treat daily activities through the operating user's communication (UIS). While previous research has read the equipment that Aids can understand, understand the UI formation, the results of agents and its private actions – especially dangerous or unrepentant – remarriage – remains. In this project, we investigate the implications of the world and the consequences of UI acids taken by AI agents. We first improving the impact rates of the mobile images of UI capabilities of a series of domain opms. After this, we conducted a Data Synensis to collect the actual UI screen and the data of the acts involved. We then use our categories of influence explaining our collected data and data turned from existing datasets in the existing UI Navigation. Our many different test languages ​​(LLMS) and diversity show how different lms can understand the impact of the mobile UI actions that may be taken by the agent. We show that taxomeny improves the facts to consult the lls of the UI's actions, but also exposed valuables in their reliable or complex fields of reliable impact.

• * Work done while in apple
• 30 University of Washington