Abstract
The rapid integration of Generative AI into Automated IT Operations (AIOps) has introduced "AI Sentinels", an autonomous agents capable of managing critical infrastructure. However, these systems introduce a novel attack surface evidenced in inference-time adversarial manipulations such as prompt injection and jailbreaking. While existing security paradigms protect network perimeters, they fail to safeguard the internal logic of AI agents, creating a research gap in runtime defense for autonomous infrastructure controllers. This study aims to develop a multi-layered, defense-in-depth architecture to neutralize these threats. The proposed system integrates three layers: an Intent Validation Engine (Layer 1) using semantic analysis, a Secure Sandbox (Layer 2) utilizing eBPF-based kernel monitoring within a digital twin, and a Static Analysis module (Layer 3) for infrastructure-as-code (IaC) compliance. Key findings indicate that while single-layer defenses achieve an Adversarial Success Rate (ASR) of 32–68%, the proposed multi-layered approach reduces the ASR to near-zero (0.2% in robust testing), maintaining an F1-score of 0.990. Despite the complexity of the pipeline, the system achieves a mean operational latency of 48.2ms on enterprise-grade hardware (NVIDIA A100). These implications suggest that runtime behavioral verification is essential for the safe deployment of LLMs in privileged environments, providing a foundational framework for resilient AIOps.