Agentic AI in E-Commerce: Security Risks of Autonomous Decision Loops in Discovery, Pricing, and Routing Systems

Abstract

Agentic artificial intelligence is becoming increasingly prevalent in large-scale e-commerce systems, where it controls critical functions including product discovery, dynamic pricing, and fulfillment routing. Unlike traditional decision support systems that need human approval before taking action, agentic architectures can act directly through platform services, creating ongoing cycles of observing, deciding, and acting with very little human involvement. While such systems reduce response latency and improve operational efficiency, they simultaneously introduce a fundamentally new category of security and integrity hazards that extend well beyond conventional application security concerns. This article examines how the extension of autonomous decision loops creates attack surfaces that traditional security frameworks fail to address adequately. The article focuses on failure modes and adversarial risks across three key commerce domains: product discovery and ranking, dynamic pricing and promotion management, and fulfillment routing optimization. By carefully examining these areas, the study finds common ways that systems can be exploited, such as manipulating feedback, hacking rewards, poisoning data, and misusing permissions by agents with too much authority. A specific threat model for agentic commerce systems is provided, along with controls for architecture and operations that focus on enforcing rules, ensuring visibility, limiting potential damage, and establishing governance. The main point made is that to protect agentic e-commerce systems, we need to prioritize decision integrity and loop stability as key security issues, which calls for new methods that work alongside, not instead of, traditional security practices.