Agentic architectures are rapidly evolving, enabling AI agents to autonomously orchestrate workflows, access heterogeneous data sources, and act on behalf of users. However, this increased autonomy amplifies a fundamental challenge: how do we ensure that agents only access—and reason over—data they are truly authorized to use? In this talk, we reflect on access control as a first-class concern in agentic systems, going beyond traditional models designed for static applications. We argue that Attribute-based Access Control provides a natural foundation for these systems, as it explicitly captures the interplay between users, agents, and execution environments. This perspective enables the enforcement of the principle of least privilege across complex, multi-agent scenarios, where the same agent may operate under different constraints depending on context.The talk concludes with open questions for the software development community, including how to make authorization transparent to agent design and how to embed secure access control mechanisms as a core building block of future agentic systems.
