Use this link to cite:
https://hdl.handle.net/2183/48265 Automatización de la mitigación de vulnerabilidades mediante OpenVAS y modelos de lenguaje
Loading...
Identifiers
Publication date
Authors
Barbeyto Torres, Daniel
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen]: La creciente cantidad de vulnerabilidades e incidentes en los equipos informáticos evidencia la necesidad de herramientas para tanto detectar como mitigar las amenazas que estas provocan. Sin embargo, aunque existen varias soluciones comerciales que realizan esta tarea, su coste es muy elevado. Por otra parte, las herramientas open source tradicionales de gestión de vulnerabilidades, como OpenVAS, son muy eficaces en la fase de detección, pero delegan su remediación en el operador humano, lo que ralentiza la respuesta ante incidentes críticos. Con el objetivo de abordar esta problemática y motivado por el auge de la inteligencia artificial, este documento propone el diseño e implementación de un sistema automatizado orientado a redes internas, que integra OpenVAS con modelos de lenguaje (LLMs) para cerrar el ciclo desde la detección de las vulnerabilidades hasta su mitigación. La herramienta desarrollada actúa como una solución centralizada que permite auditar múltiples equipos dentro de la misma red, realizando escaneos a las máquinas objetivo que generan informes con sus vulnerabilidades presentes para procesarlos y generar, de forma autónoma, scripts de seguridad personalizados para remediar los fallos detectados. La arquitectura de la herramienta usa un agente inteligente, basado en la API de OpenAI, que interpreta los resultados del escáner y genera los scripts Bash ejecutables para su posterior mitigación, mediante reglas de cortafuegos o reconfiguración de servicios, entre otros. Además del desarrollo, el documento incluye una evaluación detallada centrada en la eficacia de las mitigaciones generadas y el rendimiento del sistema en un entorno virtualizado controlado. Para ello se han realizado varias pruebas de validación tanto unitarias como masivas, probando que la herramienta puede reducir drásticamente la superficie de ataque sin ningún tipo de asistencia humana y logrando una tasa de éxito superior al 90% en la generación y ejecución de los parches de seguridad. La herramienta ha sido implementada siguiendo una metodología iterativa-incremental, permitiendo refinar el comportamiento del agente que orquesta el pipeline y asegurar la robustez de la automatización mediante revisiones continuas. Se ha definido una planificación inicial con seguimiento de hitos, y se realizó un análisis detallado de la viabilidad técnica y económica de la solución propuesta. Como resultado, se proporcionan pruebas de concepto funcional que demuestran el potencial de los modelos de lenguaje para asistir en tareas complejas de administración de sistemas y ciberseguridad, sentando las bases para futuras investigaciones en defensa de sistemas automatizados.
[Abstract]: The escalating number of vulnerabilities and incidents in computer systems underscores the need for tools capable of both detecting and mitigating the resulting threats. While several commercial solutions address this task, their cost is often prohibitive. Conversely, traditional open-source vulnerability management tools, such as OpenVAS, are highly effective in the detection phase but delegate remediation to the human operator, slowing down the response to critical incidents. To address this issue and motivated by the rise of artificial intelligence, this document proposes the design and implementation of an automated system oriented towards internal networks, which integrates OpenVAS with Large Language Models (LLMs) to close the loop from vulnerability detection to mitigation. The developed tool acts as a centralized solution allowing the audit of multiple devices within the same network, performing scans on target machines that generate reports on detected vulnerabilities and process them to autonomously generate personalized security scripts to remediate the flaws. The tool’s architecture uses an intelligent agent, based on the OpenAI API, which interprets scanner results and generates executable Bash scripts for subsequent mitigation, using firewall rules or service reconfiguration, among others. In addition to development, this document includes a detailed evaluation focused on the efficacy of generated mitigations and system’s performance in a controlled virtualized environment. For this purpose, several validation tests, both unit and massive, have been conducted, proving that the tool can drastically reduce the attack surface without any human assistance, achieving a success rate exceeding 90% in the generation and execution of security patches. The tool has been implemented following an iterative-incremental methodology, allowing the refinement of the agent orchestrating the pipeline and ensuring automation robustness through continuous reviews. An initial schedule with milestone tracking was defined, and a detailed analysis of the technical and economic feasibility of the proposed solution was performed. As a result, functional proofs of concept are provided demonstrating the potential of language models to assist in complex system administration and cybersecurity tasks, laying the groundwork for future research in automated defense systems.
[Abstract]: The escalating number of vulnerabilities and incidents in computer systems underscores the need for tools capable of both detecting and mitigating the resulting threats. While several commercial solutions address this task, their cost is often prohibitive. Conversely, traditional open-source vulnerability management tools, such as OpenVAS, are highly effective in the detection phase but delegate remediation to the human operator, slowing down the response to critical incidents. To address this issue and motivated by the rise of artificial intelligence, this document proposes the design and implementation of an automated system oriented towards internal networks, which integrates OpenVAS with Large Language Models (LLMs) to close the loop from vulnerability detection to mitigation. The developed tool acts as a centralized solution allowing the audit of multiple devices within the same network, performing scans on target machines that generate reports on detected vulnerabilities and process them to autonomously generate personalized security scripts to remediate the flaws. The tool’s architecture uses an intelligent agent, based on the OpenAI API, which interprets scanner results and generates executable Bash scripts for subsequent mitigation, using firewall rules or service reconfiguration, among others. In addition to development, this document includes a detailed evaluation focused on the efficacy of generated mitigations and system’s performance in a controlled virtualized environment. For this purpose, several validation tests, both unit and massive, have been conducted, proving that the tool can drastically reduce the attack surface without any human assistance, achieving a success rate exceeding 90% in the generation and execution of security patches. The tool has been implemented following an iterative-incremental methodology, allowing the refinement of the agent orchestrating the pipeline and ensuring automation robustness through continuous reviews. An initial schedule with milestone tracking was defined, and a detailed analysis of the technical and economic feasibility of the proposed solution was performed. As a result, functional proofs of concept are provided demonstrating the potential of language models to assist in complex system administration and cybersecurity tasks, laying the groundwork for future research in automated defense systems.
Description
Editor version
Rights
Attribution 4.0 International








