Title:

Dormant permissions, active agents: what access control auditing teaches us about agentic AI

Keynote speaker:

Simon Parkinson, Prof, Professor of Cyber Security, Manchester Metropolitan University

Simon Parkinson is a Professor of Cyber Security at Manchester Metropolitan University, where his research addresses the secure identification, authentication, and governance of people, devices, and autonomous systems. He is a member of the UK Government’s Cyber Security Advisory Board and of the Department for Science, Innovation and Technology (DSIT) College of Experts. His expertise spans digital identity, access control, behavioural biometrics, operational technology security, cyber resilience, and the application of artificial intelligence to security analysis, compliance assessment, and risk management. The common thread is turning complex access and configuration data into evidence organisations can act on. More specifically, identifying weaknesses in access control systems, detecting cyber intrusions, and preventing unauthorised access using emerging biometric technologies. Professor Parkinson has led more than £2.7 million of externally funded research and knowledge exchange, supported by EPSRC, Dstl, Innovate UK, and industry partners. He has supervised 22 research degrees to completion and has been recognised repeatedly for his supervision, winning multiple awards.

Abstract: 

Access control policies determine who can do what, and to what. In practice, they accumulate over-privilege, inconsistencies and irregularities that are difficult to identify in large, long-lived infrastructures. Historically, much of this over-entitlement went unnoticed because users rarely attempted to access resources beyond their immediate needs. Agentic Artificial Intelligence (AI) changes this. As agents are given increasingly open-ended tasks, they explore their environment, discover excess permissions and use them to achieve their goals.
In this talk, I present my research into the automated auditing of access control, covering permission modelling, misconfiguration and anomaly detection, and the importance of explanations that administrators can act upon. These lessons are then applied to agentic AI, where agents are granted permissions to invoke tools, access data and interact with operational systems, yet those permissions are rarely audited. The talk concludes with open challenges, taking the position that automated auditing should complement, and not replace, human expertise.