Logo image
Back
Privacy Risks and Leakage Pathways in Agentic and Multi-Agent AI Systems
Journal article   Open access   Peer reviewed

Privacy Risks and Leakage Pathways in Agentic and Multi-Agent AI Systems

Arash Mahboubi, Seyit Camtepe, Keyvan Ansari, Thanh Hang Bui and Hamed Aboutorab
ACM transactions on privacy and security, Just Accepted
2026
DOI: https://doi.org/10.1145/3848128
pdf
Pre-Proof698.62 kBDownloadView
Open Access CC BY V4.0

Abstract

Computing methodologies Information flow control Intelligent agents Multi-agent systems Privacy protections Security and privacy
Agentic and multi-agent AI systems are increasingly deployed as part of operational systems, introducing autonomous planning, delegation, persistent memory, tool use, and inter-agent coordination into data-driven decision making. These architectural properties challenge prevailing privacy assumptions because leakage may arise structurally from autonomy, coordination, persistence, and composition rather than only from isolated model outputs. This paper provides a structured synthesis of privacy risks and leakage pathways for agentic and multi-agent AI systems. We introduce a unifying abstraction that combines decentralized decision making with trace-based information-flow semantics, enabling reasoning about autonomy, coordination, persistent state, observability, and adversarial trace projections. Using this abstraction, we organize recurring privacy risks into five leakage pathways: autonomy-induced overexposure, coordination-driven information propagation, memory and persistent-state leakage, emergent and compositional inference, and observability and audit leakage. We further analyze why established privacy-enhancing technologies provide incomplete coverage in agentic settings, where static computation boundaries, single-agent assumptions, ephemeral state, and non-emergent behavior often fail. To ground the analysis, we implement a controlled multi-agent privacy sandbox with privacy-aware agents, tool-mediated workflows, adversarial trace projections, and LLM-backed validation. The results show that policy-checked privacy-aware execution reduces direct disclosure and adversarial inference advantage, while residual tool-mediated and semantic-hint leakage remains. The paper concludes by proposing new privacy requirements for agentic and multi-agent AI systems.

Details

Metrics

1 Record Views
Logo image