Hijacking Large Language Models: Vulnerabilities, Threat Models, Exploits, and Mitigation Strategies

Large Language Models (LLMs) are increasingly integrated into diverse applications, making their security and reliability paramount. This research addresses the importance of understanding and mitigating LLM hijacking – malicious manipulation aimed at controlling or corrupting model outputs. Current research often focuses on specific attack vectors, but a holistic understanding, standardized threat modeling, and systematic evaluation of defenses remain a research gap. This paper's research objectives are to: (1) define and categorize LLM hijacking techniques; (2) conduct a comprehensive review of existing vulnerabilities; (3) develop a formal threat model adapted for LLMs; (4) simulate a representative hijacking technique (prompt injection) to demonstrate feasibility; (5) propose and discuss mitigation strategies; and (6) analyze the ethical and policy implications. The methodology combines literature synthesis, theoretical threat modeling (adapting STRIDE), and computational experimentation using a simulated prompt injection attack on a pre-trained LLM (GPT-2 architecture via Hugging Face's `transformers`). Key findings demonstrate the ease with which simple prompt injection can override intended instructions in a standard LLM, highlighting significant vulnerabilities. The simulation confirms that carefully crafted adversarial prompts can reliably hijack model output. Implications of this research underscore the urgent need for robust defense mechanisms, standardized security evaluation protocols for LLMs, and proactive policy development to govern their safe deployment, mitigating risks such as misinformation generation, data exfiltration, and unauthorized system interactions.