Securing Agentic AI: Threat Modeling and Risk Analysis for Network Monitoring Agentic AI System
Pallavi Zambare 1 {}^{\textbf{1}}, Member, IEEE, Venkata Nikhil Thanikella 1 {}^{\textbf{1}}, and Ying Liu 1 {}^{\textbf{1}}, Member, IEEE · 2025
Read Securing Agentic AI: Threat Modeling and Risk Analysis for Network Monitoring Agentic AI System by Pallavi Zambare 1 {}^{\textbf{1}}, Member, IEEE, Venkata Nikhil Thanikella 1 {}^{\textbf{1}}, and Ying Liu 1 {}^{\textbf{1}}, Member, IEEE (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-10043.txt.
Introduction
Pallavi Zambare, Venkata Nikhil Thanikella, and Ying Liu probe the security of agentic AI used in network monitoring, in this 2025 paper. Agents built on large language models can watch traffic and act on their own, which also opens new attack surfaces. The team applied a seven-layer threat modeling frame called MAESTRO to a working prototype.
They built the agent with Python and LangChain, fed by live telemetry. Two attacks were confirmed in practice. Replaying traffic flooded the agent and delayed its updates, a denial of service. Tampering with the agent's stored log poisoned its memory, bending later decisions and raising compute load. The authors propose layered defenses: isolate memory, validate plans, and watch adaptation logic in real time. They argue such checks are needed before agents guard real networks.
Catalog record
| CREATOR | Pallavi Zambare 1 {}^{\textbf{1}}, Member, IEEE, Venkata Nikhil Thanikella 1 {}^{\textbf{1}}, and Ying Liu 1 {}^{\textbf{1}}, Member, IEEE |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-10043
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Securing Agentic AI: Threat Modeling and Risk Analysis for Network Monitoring Agentic AI System Pallavi Zambare 1 {}^{\textbf{1}}, Member, IEEE, Venkata Nikhil Thanikella 1 {}^{\textbf{1}}, and Ying Liu 1 {}^{\textbf{1}}, Member, IEEE Abstract When combining Large Language Models (LLMs) with autonomous agents, used in network monitoring and decision-making systems, this will create serious security issues. In this research, the MAESTRO framework consisting of the seven layers threat modeling architecture in the system was used to expose, evaluate, and eliminate vulnerabilities of agentic AI. The prototype agent system was constructed and implemented, using Python, LangChain, and telemetry in WebSockets, and deployed with inference, memory, parameter tuning, and anomaly detection modules. Two practical threat cases were confirmed as follows: (i) resource denial of service by traffic replay denial-of-service, and (ii) memory poisoning by tampering with the historical log file maintained by the agent. These situations resulted in measurable levels of performance degradation, i.e. telemetry updates were delayed, and computational loads were increased, as a result of poor system adaptations. It was suggested to use a multilayered defense-in-depth approach with memory isolation, validation of planners and anomaly response systems in real-time. These findings verify that MAESTRO is viable in operational threat mapping, prospective risk scoring, and the basis of the resilient system design. The authors bring attention to the importance of the enforcement of memory integrity, paying attention to the adaptation logic monitoring, and cross-layer communication protection that guarantee the agentic AI reliability in adversarial settings. {IEEEkeywords} Autonomous Agents, MAESTRO Framework, Threat Modeling, Network Monitoring, LLM Security, Risk Assessment. † † corresponding: Corresponding author:Ying Liu (email:[email protected]). 1 INTRODUCTION 1.1 OVERVIEW OF AGENTIC AI AND NEED OF AUTONOMOUS AGENTS IN NETWORK MONITORING An agentic AI can be described as autonomous memory-based systems that engage in planning, relying, and communication with external tools in a behaviorally self-directed way. Compared to the reactive patterns of AI models that work on the basis of a single prediction, the agentic systems follow the perception-reasoning-acting loop. This allows them to make choices and aim at achieving objectives and changing with time. They are powered either by large language models (LLMs) or multimodal foundation models with the ability to reason in context, operate with long-term horizons, and co-ordinate multimodal actions that are not merely short-term responsive [ 1 ] , [ 2 ] . Agents agentic architectures liberate agents stateful and compositional behaviors and actions, and planning beyond stimulus-response behavior, so are more suitable to dynamic environments where rule-based behavior cannot be adapted dynamically. An example is in network monitoring, where the conventional tools and systems such as static intrusion detection system (IDS) today as well as the threshold monitors are inadequate toward identifying new threats or zero-day threats [ 3 ] . The gap is closed in agentic AI by the ability to conduct on-the-fly telemetry analysis, context sensitive alerts, and auto-tunable parameters. A machine trained on an LLM could also make the difference between benign and malicious traffic influxes and respond, with hardly any guidance on the human side. Such agents can change constantly through feedback loops, and thus provide quick, adaptive solutions in cybersecurity situations. They improve resiliency through proactive mitigation of threats, cross-correlation of trends and detection gaps, thus the need to use agentic AI in the current dynamic threat environment [ 4 ] . 1.2 GROWING DEPLOYMENT OF LLM-AUGMENTED AGENTS IN CYBERSECURITY Large language models (LLMs) have allowed large scale deployments in cybersecurity to replace traditional, human-manned coordinating effort with semi-autonomous entities that perform threat, anomaly analysis, and response generation in real time. Based on the generative and contextual reasoning capabilities of LLMs, these agents can execute sophisticated work that typically needs a human background to conduct, including log parsing, incident summarization, and policy refinement. They are particularly suitable in the dynamic environment of such applications as network defense since they can be adapted to unstructured information and can learn through feedback. Some of the areas which LLMs are applicable include intrusion detection, phishing classification, alert triage, and vulnerability management. As an example, our model applied to a SIEM system using a transformer-based prediction correctly classified a substantial number of false positives with the help of correlating alert history with metadata context in the past [ 5 ] . Figure 1: A generalized architecture of a single-agent Agentic AI system Threat hunting as well as incident response also relies heavily on LLM-powered agents to help them extract indicators of compromise (IOCs) posted in live telemetry and execute automated responses in situations like firewalls updates [ 6 ] . These agents support workflow orchestration without scalable servers and distributed threat mitigation, and visibility and overall effectiveness in cloud-native infrastructures [ 7 ] . On the other hand, trust and control issues have been offered by such challenges as hallucinations, goal disparities, adversarial manipulation (e.g., memory poisoning), and partial observability. Their increase in decision making autonomy leads to an appetite among LLM agents to require more advanced threat modeling systems and architectural defenses in order to be deployed safely and in an auditable way. 1.3 CHALLENGES: AUTONOMY, EXTERNAL TOOL USE, REASONING LOGIC — NEW THREAT CLASSES. EXISTING FRAMEWORKS (E.G., STRIDE, PASTA) ARE INADEQUATE. New security risks are presented by agentic AI systems, particularly large language models (LLMs) driven systems, predicated on an expansiveness that traditional rule-based or supervised AI could not support. Such agents have advanced features including autonomous reasoning, persistent memory, adaptive planning, and external tools invocation, enlarging their scale of functionality considerably, and, respectively, the surface of attack. The threats arising include goal misalignment, memory poisoning, and multi-stage reasoning hijacks because of the changing behavior of agents in the dynamic environment. Such frameworks as STRIDE and PASTA do not have the ability to deal with such complex interactions, as all of the architectures are modeled as static, and are not deep in modeling semantic or emergent behavior. Overcoming these limitations, this work implements the MAESTRO framework that divides the architecture of an agent into 7 interrelated layers so that the threat localization at different layers and the risk scoring based on a specific AI are possible. The questions the research produces respond to are the following: 1. How do threats emerge in autonomous agents? 2. How can they be mapped and evaluated efficiently using MAESTRO ? 3. What mitigation strategies are the best? It gives the contextual MAESTRO mapping, threat taxonomy, a risk scoring model based on severity, and a real-world validation in the scenario of a deployed LLM-based agent in a live telemetry monitoring attitude. 2 Background and Related Work 2.1 Existing Security Frameworks The established models of security, such as STRIDE, PASTA, OWASP Top 10 have long been utilized in secure software development and insufficiently accommodate the agentic AI systems that feature autonomy and dynamic reasoning. To show how to apply STRIDE to structured systems such as infotainment, Das et al. [ 8 ] found 34 threats. Tete et al. [ 9 ] and Gulen et al. [ 10 ] have also generalized STRIDE to LLM-based applications and have typified concerns such as injection and privilege abuse. Nevertheless, STRIDE does not model emergent behavior, cognitive reasoning of AI agents very well. Juuso [ 11 ] pointed out that the threats associated with invoking