Skip to content
Conference Open access

Enterprise Agentic AI Infrastructure: A Zero Trust Runtime Security Framework for Secure Autonomous Multi-Agent Systems

Oct 2026 · 2026 IEEE Conference on Generative AI for Secure Systems (GAISS) · pp. 1-8 · 0 citations · 36 references

Abstract

Enterprises are adopting large language model (LLM)-based agentic artificial intelligence (AI) at a rapid pace: autonomous agents now invoke tools, execute multi-step workflows, retain long-term memory, and coordinate with peer agents under limited human supervision. Enterprise deployment of autonomous agents fundamentally expands the attack surface beyond the foundation model: prompt injection, tool abuse, memory poisoning, agent impersonation, privilege escalation, and compromised Model Context Protocol (MCP) servers are now practical, demonstrated threats. Existing defenses concentrate on model-level guardrails while leaving the surrounding runtime largely unsecured — a runtime security gap. We addressed this gap with a vendor-neutral Zero Trust Runtime Security Framework that extends Zero Trust principles across the agent lifecycle: identity issuance, prompt-level threat detection, policy-as-code enforcement, memory integrity verification, least-privilege tool authorization, and a secure MCP gateway. The framework introduces a runtime risk-scoring engine fusing prompt, tool, memory, identity, context, and reputation signals into allow/approve/block/quarantine decisions, paired with an autonomous incident-response loop for detection, correlation, containment, and recovery. We developed a threat taxonomy for enterprise multi-agent deployments. The evaluation methodology we proposed uses attack success rate, detection rate, false-positive rate, and mean time to response. The contribution is architectural rather than a specific commercial implementation; empirical, production-scale validation is identified as future work.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.