Real-Time AI Agent Governance: Per-Agent Permission Controls, Behavioral Anomaly Detection, and Policy Enforcement for Desktop AI Tools

Read full text Summarise with VIRAT
Authors: Anil Kumar Sharma
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Desktop AI tools — ChatGPT Desktop, Claude Desktop, GitHub Copilot, Cursor, Grammarly, and an expanding ecosystem of autonomous agents — have become first-class participants in personal and professional computing workflows. Unlike traditional applications, these tools are granted broad, often unrestricted operating system access: they read file systems, issue network requests, interact with clipboard contents, and in some configurations execute arbitrary code. As of January 2026, no system existed that provided users or organizations with (a) real-time visibility into what AI agents are doing at the OS level, (b) granular per-agent permission controls analogous to mobile app sandboxing, (c) declarative policy enforcement scoped to AI agent identity, or (d) behavioral anomaly detection that establishes per-tool baselines and alerts on deviation. This paper presents the design of AnkrShield AI Agent Governance, a desktop security layer that addresses all four gaps. The architecture introduces a tri-signal AI agent discovery mechanism (process signatures, network traffic fingerprinting, file system access correlation), a per-agent permission model decoupled from OS-level user permissions, a declarative policy engine with a novel AI-principal policy language, a behavioral baseline system using lightweight statistical profiling, and a hash-chained tamper-evident audit trail exportable for compliance under India's Digital Personal Data Protection Act 2023 and SOC 2. Crucially, the system operates entirely locally — it governs AI agents without becoming an agent itself.

Fields of Study & Subjects

Access Control and Trust Security and Verification in Computing Network Security and Intrusion Detection

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN
Document type
Preprint