Security Vulnerabilities in Large Language Models: Bibliometric Analysis

Abstract

The sudden rise of large language models (LLMs) has further increased the concern for their security related to weaknesses such as adversarial attacks related to prompting, jailbreaking attacks, and data leakage. However, a comprehensive understanding related to the developments in this area of research has been restricted despite numerous publications being generated along these lines. The current research encompasses a bibliometric analysis related to security weaknesses associated with LLMs within 410 peer-reviewed publications on Scopus and Web of Science from 2021-2025. Although the eligibility window starts in the year 2021, there are no publications that are considered to be eligible for that year. There are a few early-access publications and preprints from the year 2026 that are included if they are indexed in that year, as this is a rapidly growing area of literature. The study uses the Bibliometrix package in the R programming language to integrate performance measures with concept and theme mapping methods such as keyword co-occurrence graphs, theme development measures, trend topic analysis, and three-way plots that relate sources to authors and keywords. Findings show an explosion in research work post-2023 driven by the rapid development in prompt attacks and jailbreak methods. Conceptual analysis also reveals that adversarial threats and jailbreak threats are central to LLM security studies, which are largely tied to conceptual probes on foundational models. Data privacy risks such as data leakage and membership inference attacks emerge as technologically rich but relatively niche topics, while system security perspectives are gradually gaining prominence. Also, analysis shows that there is a paradigm shift from foundational AI studies and NLP towards security applications and safety regulation. In conclusion, this work illustrates the fast-paced maturity as well as the fragmentation of the LLM security body of work. By integrating the literature that has been conducted to date, the results above have aided in the foundation of a resource that will enable the building of more comprehensive models for securing large language models.