Automated CVE Triage: Ai-Agent Framework for Scalable Vulnerability Triaging and Security Automation

As vulnerability disclosure platforms scale, the manual triage of security reports has become a significant operational burden. Government and industry systems such as the Common Vulnerabilities and Exposures (CVE) database, HackerOne, Bugcrowd, and others rely on teams of analysts to validate vulnerability submissions, reproduce proof-of-concept (PoC) exploits, identify affected software versions, and assign Common Vulnerability Scoring System (CVSS) metrics. This process demands substantial labor and funding, contributing to disclosure backlogs and delayed remediation. In this paper, we propose an AI-agent-based triage framework that automates core tasks traditionally performed by human analysts, which has some limitations on the current model driven centrally by MITRE. Our system integrates large language models (LLMs), implemented as small agentic AI systems trained for specific tasks, with sandboxed terminal environments to autonomously interpret reports, execute PoCs in isolated settings, and generate structured outputs suitable for CVE publication or platform response workflows. Initial results indicate that the framework can significantly reduce triage time and operational costs, offering a scalable and reproducible alternative to manual processing.9