LLM based prompt injection detection accuracy and its relationship to the context size

Prudnikovas, Aivaras (2025) LLM based prompt injection detection accuracy and its relationship to the context size. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (1MB) | Preview

Abstract

LLMs, widely used in various applications, are vulnerable to malicious prompt injections that can manipulate their outputs, potentially leading to security breaches. Prompt injection attacks exploit the model’s design to follow instructions allowing attackers to change the outputs by injecting malicious instructions. This research evaluates how the prompt size influences the accuracy of the prompt injection attacks. It also evaluates the accuracy of two detection techniques (naïve and reinforced) on GPT-4.1, Ministral-3b, Mistral-large, Llama-4-Maverick, Gemma3. A set of more than 1000 injection prompts are wrapped in varying size content to increase context (no increase, 1000 tokens, 5000 tokens, 10000 tokens, 50000 tokens) and detection performance is measured for each configuration. The results show the detection accuracy first increases when context size is inflated with 1000 tokens of HTML on all LLMs and detections, then it varies without a clear relationship. GPT-4.1 outperforms other models peaking at 98.8% accuracy and a reinforced detection outperforms the naïve one.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Mustafa, Raza Ul UNSPECIFIED
Uncontrolled Keywords:	prompt; injection; context size; accuracy; LLM; detection
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence P Language and Literature > P Philology. Linguistics > Computational linguistics. Natural language processing Q Science > QA Mathematics > Computer software > Computer Security T Technology > T Technology (General) > Information Technology > Computer software > Computer Security
Divisions:	School of Computing > Master of Science in Cyber Security
Depositing User:	Ciara O'Brien
Date Deposited:	19 Aug 2026 15:05
Last Modified:	19 Aug 2026 15:05
URI:	https://norma.ncirl.ie/id/eprint/9550

Actions (login required)

View Item