Secure Authentication Design For AI Agents

Abstract

In the financial industry, artificial intelligence (AI) agents are increasingly adopted in order to drive higher productivity and financial performance. These solutions require access to critical enterprise systems like ERPs, trading platforms or other solutions where they need to authenticate and execute actions on behalf of their users. This is brings specific security challenges on how to reliably authenticate the agents to these critical systems. In this paper we will explore the common anti-patterns on how to design the authentication mechanism for agents in function calling use cases. Additionally we will provide the best solution for implementing the authentication and explore two alternative security solutions depending on the capabilities of the external system. Finally we will provide an example architecture of using the MCP protocol while authenticating the agent.

Keywords: AI Agents, Security in AI, Function Calling, MCP

DOI: 10.54941/ahfe1006933