A framework for cyber threat intelligence sharing focused on AI vulnerabilities

Abstract

A new area of cybersecurity concern has emerged as a result of the increasing application of artificial intelligence (AI) into autonomous systems, healthcare, and critical infrastructure. The specificity needed to identify and address AI-specific vulnerabilities, which can appear at different phases of the AI lifecycle (development, training, and deployment \& use), is frequently lacking in traditional Cyber Threat Intelligence (CTI) systems and taxonomies. In order to meet the pressing need for a customized approach to threat intelligence in AI systems, this thesis investigates the intersection of CTI sharing and AI vulnerability categorization. A scalable, web-based framework that facilitates both automatic CVE ingestion and community-driven reporting is used to construct a novel taxonomy based on the paper, "A Comprehensive Artificial Intelligence Vulnerability Taxonomy" (AIVT). The platform offers a centralized system for recording hazards particular to AI by integrating a structured classification pipeline using Large Language Models (LLMs), secure Application Programming Interface (API) and PostgreSQL-based storage. In order to find gaps in taxonomy, lifecycle awareness, and attribute-level analysis, the study assesses current platforms including AI Vulnerability Database (AVID), AI Incident Database (AIID), and Massachusetts Institute of Technology Research and Engineering Adversarial Threat Landscape for Artificial-Intelligence Systems (MITRE ATLAS). This research offers a structured and practical CTI-sharing strategy to improve the security and reliability of AI systems by putting forward and executing a three-stage taxonomy that covers the AI phase, compromised attributes, and exploitation impacts.