Agents Without Anchors: A Taxonomy of Goal Drift in Intelligent Reality
Year: 2026
Type: conference-paper
Abstract: Agentic AI systems are becoming foundational to Intelligent Reality environments, coordinating digital twins, managing cyber-physical systems, and mediating metaverse interactions. As these systems operate autonomously over extended deployments, they are susceptible to goal drift - systematic divergence between operational behavior and sanctioned objectives, occurring without explicit reprogramming and without triggering conventional monitoring alerts. This paper presents a taxonomy of five goal drift types in production agentic IR systems: Objective, Scope, Context, Constraint, and Temporal Drift. Each type is grounded in documented mechanisms from the AI alignment and concept drift literature and illustrated through representative IR deployment scenarios. We analyze each drift type through dual lenses as a security vulnerability and as an ethical accountability failure and derive implications for IR system design and governance. Our taxonomy provides a foundation for drift-aware monitoring, authorization frameworks, and governance standards in IR deployments where physical-world coupling amplifies the consequence of undetected behavioral divergence.
Authors: Rohan Vardhan, Shrikant Sonparote
Institutions: Meta (United States), Lowe's (United States)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202128692