Authors: Sonal Sinha
Abstract: Artificial Intelligence (AI) has a dual relationship with cybersecurity: it is both a powerful defensive technology and, increasingly, a high-value attack target in its own right. Machine learning, deep learning, foundation models, and large language models (LLMs) now underpin intrusion detection, malware analysis, phishing identification, and autonomous incident response. At the same time, adversaries exploit AI for automated reconnaissance, AI-generated phishing, and attacks aimed directly at AI models — adversarial evasion, data poisoning, model extraction, prompt injection, jailbreaking, and AI supply-chain compromise. This paper condenses a comprehensive survey into a conference-length treatment: it traces the evolution of AI in cybersecurity, presents a unified taxonomy of attacks across the AI lifecycle, reviews key defense mechanisms (explainable AI, federated learning, differential privacy, guardrails, zero trust), summarizes major governance frameworks (NIST AI RMF, ISO/IEC 42001, EU AI Act, OWASP LLM Top 10), and outlines open research challenges for building trustworthy, resilient AI-driven security systems.
