Large Language Models (LLMs) have become integral to numerous digital systems, yet their susceptibility to adversarial misuse remains a critical concern. This thesis investigates three major categories of LLM vulnerabilities—prompt injection, jailbreaking, and hallucination across three open-source models: Vicuna-7B, Falcon-RW-1B, and OpenChat-3.5. A custom evaluation framework was developed using Python and Hugging Face libraries to systematically test and compare model behavior against adversarial prompts sourced from public datasets. The results reveal significant variation in model resilience. Falcon-RW-1B, despite being the smallest model, consistently demonstrated the highest resistance to adversarial inputs, suggesting that model size alone does not determine security. In contrast, OpenChat-3.5, while linguistically powerful, exhibited the highest vulnerability rates across all test types. Vicuna- 7B displayed moderate susceptibility, particularly to hallucination and contextual overrides. The key conclusion is that instruction tuning focused on helpfulness can unintentionally weaken a model’s ability to reject malicious prompts. The research further demonstrates that lightweight, quantized models can be safely evaluated in constrained environments without compromising the integrity of security analysis. This thesis contributes a reproducible testing pipeline and offers evidence-based recommendations for improving the robustness of LLMs in future deployments.