Date of Publication : August 2026

Abstract: Large language model (LLM) applications are moving beyond passive text generation toward agentic systems that retrieve external knowledge, use tools, call APIs, and interact with enterprise data within the same request lifecycle. This shift introduces runtime security concerns that conventional infrastructure monitoring, offline safety testing, and standalone content moderation do not address well. AgentGuard AI is developed as a unified runtime governance framework for these environments. It brings together three mechanisms: the Secure Runtime Monitoring Protocol (SRMP), which records infrastructure telemetry, model behaviour, agent actions, retrieval events, security alerts, and audit information in a common event structure; Agentic and Accelerator-Aware Risk Scoring (AARS), which produces an interpretable composite risk score from six dimensions; and the Risk-Adaptive Scheduling Strategy (RASS), which translates the resulting risk profile into an operational response such as allowing the request, throttling it, routing it for human review, or blocking it. The prototype was evaluated on 300 controlled requests representing normal, attack, long-context, and tool-intensive workloads. Attack detection increased from 28.3% with the baseline logging configuration to 85.0% with AgentGuard AI, a relative improvement of about 200%. Audit coverage reached 100%. A lightweight threat classifier based on TF-IDF and Logistic Regression achieved 97.8% accuracy and a 97.8% F1 score under 5-fold cross-validation. Monitoring added approximately 8-12% runtime overhead. Taken together, the results show that combining event capture, multidimensional risk assessment, and deterministic runtime controls can strengthen the security and auditability of LLM-agent deployments without sacrificing interpretability or practical deployability.

Reference :