Hossain, Muhammad IqbalPasha, Md. SalmanAhsan, KM AbrarNodi, RummanMaliha, ShawanaSiddiquee, Md. Eousuf2026-01-062026-01-0620252025-10ID 24141081ID 24141119ID 23241082ID 22101117ID 24141082http://hdl.handle.net/10361/27403Cataloged from PDF version of thesis.Includes bibliographical references (pages 52-54).This thesis is submitted in partial fulfillment of the requirements for the degree of Bachelor of Science in Computer Science, 2025.As organizations continue to accumulate more data in the digital platforms, it becomes difficult to safeguard sensitive data. The authorization and authentication factors restrict the use of critical systems, but traditional IAM can’t scale with the new breed of cyber-threats like credential theft, phishing and AI attacks. The motive of this paper is in building security systems more resilient and intelligent to deal with any kind of malicious attacks and generating decisions with the application of Large Language Models (LLMs) augmented with advanced AI driven techniques. The project was initially focused on the specified access control factors of Identity and Access Management (IAM) and evaluating policies with providing security scores connected to web interfaces to analyze vulnerable factors beforehand. This was incorporated to a hybrid AI architecture consisting of a small BERT-tiny model optimized to detect security anomalies quickly and larger transformer based models LLMs (Mistral-7B and Gemma3-270M) that can be deployed to explain problems in fine detail and generate remediation strategies that can be executed. Experiments on larger actual datasets showed BERT-tiny achieved a remarkable accuracy of 90.12% for detection and 82.45% for malicious type differentiation. The focus applied on hyperparameter tuning, multi layer approach optimization, and class imbalance adjustments ensured robustness and generalization. Although certain limitations remain, especially in aspects of response latencies and the computational overhead, the present work is a step towards demonstrating the radical potential of an LLM approach in building systems that think faster, explain better, and adapt smarter to emerging digital threats.63 pagesenBRAC University theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission.Large language modelsArtificial intelligenceBERT-tinySecurity evaluationHyperparameter tuningMistral-7BGemma3CybersecurityCyberspace--Security measures--Evaluation.Computer networks--Security measures.Computer security.Information networks--Security measures.Artificial intelligence--Security measuresThesis