Architectural Model for Controlling Autonomous AI Agents as Insider Threats in Information Security Systems
Year: 2026
Type: article
Abstract: Organizations are increasingly relying on autonomous agents based on artificial intelligence (AI) to execute everyday business processes.These systems, referred as "Agentic AI", have access to sensitive data and information systems, operating without constant human oversight.This article proposes viewing such agents as a new class of insider threat.It analyzes key attack vectors against these systems, specifically: instruction injection, excessive autonomy, tool misuse, and uncontrolled data exfiltration.The proposed approach for developing information security systems is based on a conceptual model and a UML class diagram, augmented with a new "autonomous agent control" component and an "Action Gateway" class.The latter assesses the risk associated with each agent action, determining whether it should be permitted, escalated to a human for decision making, or blocked.The framework is demonstrated through three attack scenarios within a financial organization, illustrating how the proposed mechanisms can thwart attacks that bypass other security controls in this specific configuration.The framework is compared with similar approaches for controlling agent actions (such as GuardAgent and CaMeL); security is ensured through integration with the organization's existing security architecture, as well as by routing medium-risk actions to a designated responsible individual.
Author: Ivan Gaidarski
Institutions: Bulgarian Academy of Sciences, Institute of Robotics
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7223939156