Virginia Tech® home

A Comprehensive Assessment Tool for Prompt Injection Attacks in Large Language Models (LLMs)

Research Paper Showcase 2026

Abstract

A prompt injection attack is a type of cyberattack that targets Large Language Models (LLMs), where attackers embed malicious instructions into an LLM with the intent of manipulating the model to the attacker's desire. Due to the relevance, growth, and demand of LLMs within small to medium-sized businesses, it is essential that they have access to a tool that can comprehensively assess model vulnerabilities. Based on an established framework for Prompt Injection attacks, this paper introduces a tool that can simulate types of injection attacks across different widely used LLM models and versions, test different defenses against these attacks, and measure the performance and how vulnerable the LLM's responses are. Additionally, this paper incorporates cost analysis, CPU utilization metrics, and a prompt hardening methodology for LLM personas. The results of these experiments provide deeper insights into testing and a comparative analysis against commonly used LLMs today.


Authors

  • Ailene Dao, undergraduate student (Bachelor's to Accelerated Master's, digital forensics concentration), cybersecurity engineering, George Mason University
  • Richard VanGorder, research assistant, cybersecurity engineering, George Mason University
  • Mohamed Gebril, associate professor, cybersecurity engineering, George Mason University
  • Sherif Abdelhamid, assistant professor, computer and information sciences, Virginia Military Institute

Publication

  • Venue: 2026 IEEE 5th International Conference on AI in Cybersecurity (ICAIC)
  • Date: Feb. 18-20, 2026

Related Papers

BOCLOAK: Optimal Transport-Guided Adversarial Attacks on Graph Neural Network-Based Bot Detection

A Comprehensive Assessment Tool for Prompt Injection Attacks in Large Language Models (LLMs)

An Empirical Study of LLM Serving in Confidential GPUs

FedHusky: Accelerating Hybrid Federated Learning with Client Hopping

MoltGraph: A Longitudinal Temporal Graph Dataset of Moltbook for Coordinated-Agent Detection

OmniVul: A Holistic, Multi-Turn Conversational Benchmark for LLM-Based Vulnerability Assessment

StealthInk: A Multi-bit and Stealthy Watermark for Large Language Models

Towards AI-Driven Human-Machine Co-Teaming for Adaptive and Agile Cyber Security Operation Centers