Abstract
<title>Abstract</title> <p>Large Language Model (LLM) security research is growing faster than the evaluation infrastructure needed to validate its outputs. This study provides the first integrated bibliometric mapping of defense mechanisms against adversarial attacks on LLMs, analyzing 137 Q1 journal articles published between 2024 and 2026 across 19 Scopus-indexed sources. Scientific production grew at 78.38% annually, signaling a structurally new field.Using Bibliometrix 4.1 and VOSviewer, the study applied Callon's strategic centrality-density diagrams, Multiple Correspondence Analysis (MCA), and thematic evolution analysis. `Adversarial machine learning' (23 keyword occurrences) and `contrastive learning' (20) emerged as motor themes. Thematic evolution revealed a shift toward `vulnerability detection' (weighted inclusion index = 0.71) and `benchmarking' (0.27). MCA accounted for 53.40% of conceptual variance, separating security-oriented terminology from NLP and training approaches.Five interdependent knowledge gaps were identified: adaptive scalability, benchmark standardization, emerging vulnerabilities, cultural-ethical integration, and detection automation. The gap between publication growth and evaluation infrastructure poses a governance risk for organizations deploying LLMs in high-stakes environments and constrains regulatory compliance timelines.</p>