Reproduce “A Vision for Access Control in LLM-based Agent Systems” on OpenResearch

About this paper

Xinfeng Li, Dong Huang, Jie Li, Hongyi Cai, Zhenhong Zhou, Wei Dong, XiaoFeng Wang, Yang Liu

Researchers from Nanyang Technological University and National University of Singapore introduce Agent Access Control (AAC), a conceptual framework that redefines access control for LLM-based agent systems as an intrinsic cognitive capability focused on information governance. The framework employs multi-dimensional contextual evaluation and adaptive response formulation, underpinned by a dedicated neuro-symbolic reasoning engine, to dynamically manage information flow rather than relying on static permissions.

Key insights

• Entrusting the primary LLM with its own security responsibility is a design flaw due to its susceptibility to manipulation, necessitating a separate, dedicated core reasoning engine for access control based on separation of concerns.
• The core reasoning engine for AAC should be a compact, verifiable neuro-symbolic reasoner, combining contextual understanding from neural components with robust, auditable enforcement of formal policy logic.
• Access control should be viewed as a continuous flow of information that is dynamically governed, rather than a static gatekeeping function, enabling contextually grounded, end-to-end response generation.

Summary by alphaXiv.