Year: 2025
Type: dissertation
Abstract: The rapid advancement of Large Language Models (LLMs) has powered transformative applications in many fields. However, their widespread adoption introduces significant security risks depending on deployment strategies. This study conducts a comparative analysis of security vulnerabilities in publicly accessible (GPT-4o) and privately hosted (LLaMA 3.2) LLM deployments, focusing on three critical attack vectors: jailbreak attacks, prompt injection, and data poisoning. Using a controlled case study methodology, adversarial scenarios were simulated to generate and categorize model outputs, with qualitative content analysis applied to analyze vulnerability severity and behavioral patterns. Findings reveal a critical asymmetry: public LLMs demonstrate strong resilience to direct prompt injection but exhibit high susceptibility to indirect jailbreaks, where safety disclaimers fail to prevent malicious content generation. Conversely, private LLM shows inconsistent jailbreak resistance. However, it is severely vulnerable to data poisoning and delayed detection of direct prompt injections, primarily due to static configurations and a lack of runtime safeguards. The findings of this study highlight the importance of developing tailored security measures that can adapt to targeted threats while maintaining the performance and utility of public and private LLMs. Additionally, future work should investigate hybrid attacks and adversarial inputs, while also promoting domain-specific benchmarks to ensure the responsible use of LLMs in critical settings.
Author: Wang, Zixi
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7112236095