The Cybersecurity Risks of Generative AI: Attack and Defense Perspectives

Аннотация

Background. The rapid, widespread adoption of generative artificial intelligence (GenAI) large language models (LLMs), diffusion-based image generators, and multimodal systems has created a dual-use cybersecurity problem: the same generative capabilities that increase productivity can be repurposed by attackers, while generative systems themselves expose new, AI-specific attack surfaces. This study addresses the lack of an integrated framework connecting offensive capability to defensive maturity across the GenAI lifecycle. Methods. A structured narrative literature review was conducted across IEEE Xplore, the ACM Digital Library, arXiv, and Google Scholar, screening an initial pool of approximately sixty sources published between 2014 and 2025 and retaining twenty-three that described a distinct attack technique, defensive mechanism, or governance framework applicable to generative-AI systems. Each source was coded by lifecycle stage and offensive or defensive orientation, and a qualitative Low/Medium/High maturity rating was assigned to each resulting threat category. Results. The review yields a five-category taxonomy AI-enabled offense, prompt-based manipulation, training-data extraction and privacy leakage, data poisoning and supply-chain manipulation, and adversarial evasion of AI-based detectors mapped to representative techniques and primary defenses. Defense maturity is uneven: AI-enabled offense and data-poisoning categories are rated Low to Low-Medium, while privacy-leakage and adversarial-evasion categories are rated Medium, with no category rated High. Conclusions. The main contribution is an integrated, lifecycle-based attack-defense taxonomy showing that defensive investment is concentrated at the deployment stage while earlier, data- and training-stage risks remain comparatively under-defended, implying that organizations deploying GenAI require lifecycle-wide, layered security governance rather than guardrails confined to model output.