Systematically Analysing Prompt Injection Vulnerabilities in Diverse LLM Architectures

Αποθηκεύτηκε σε:
Λεπτομέρειες βιβλιογραφικής εγγραφής
Εκδόθηκε σε:International Conference on Cyber Warfare and Security (Mar 2025), p. 142
Κύριος συγγραφέας: Heverin, Thomas
Άλλοι συγγραφείς: Benjamin, Victoria, Braca, Emily, Carter, Israel, Kanchwala, Hafsa, Khojasteh, Nava, Landow, Charly, Luo, Yi, Ma, Caroline, Magarelli, Anna, Mirin, Rachel, Moyer, Avery, Simpson, Kayla, Skawinski, Amelia
Έκδοση:
Academic Conferences International Limited
Θέματα:
Διαθέσιμο Online:Citation/Abstract
Full Text
Full Text - PDF
Ετικέτες: Προσθήκη ετικέτας
Δεν υπάρχουν, Καταχωρήστε ετικέτα πρώτοι!

MARC

LEADER 00000nab a2200000uu 4500
001 3202190576
003 UK-CbPIL
035 |a 3202190576 
045 2 |b d20250301  |b d20250331 
084 |a 142229  |2 nlm 
100 1 |a Heverin, Thomas 
245 1 |a Systematically Analysing Prompt Injection Vulnerabilities in Diverse LLM Architectures 
260 |b Academic Conferences International Limited  |c Mar 2025 
513 |a Conference Proceedings 
520 3 |a This paper presents an exploratory systematic analysis of prompt injection vulnerabilities across 36 diverse large language models (LLMs), revealing significant security concerns in these widely adopted AI tools. Prompt injection attacks, which involve crafting inputs to manipulate LLM outputs, pose risks such as unauthorized access, data leaks, and misinformation. Through 144 tests with four tailored prompt injections, we found that 56% of attempts successfully bypassed LLM safeguards, with vulnerability rates ranging from 53% to 61% across different prompt designs. Notably, 28% of tested LLMs were susceptible to all four prompts, indicating a critical lack of robustness. Our findings show that model size and architecture significantly influence susceptibility, with smaller models generally more prone to attacks. Statistical methods, including random forest feature analysis and logistic regression, revealed that model parameters play a primary role in vulnerability, though LLM type also contributes. Clustering analysis further identified distinct vulnerability profiles based on model configuration, underscoring the need for multi-faceted defence strategies. The study's implications are broad, particularly for sectors integrating LLMs into sensitive applications. Our results align with OWASP and MITREs security frameworks, highlighting the urgency for proactive measures, such as human oversight and trust boundaries, to protect against prompt injection risks. Future research should explore multilingual prompt injections and multi-step attack defences to enhance the resilience of LLMs in complex, real-world environments. This work contributes valuable insights into LLM vulnerabilities, aiming to advance the field toward safer AI deployments. 
653 |a Statistical methods 
653 |a False information 
653 |a Cluster analysis 
653 |a Large language models 
653 |a Security 
653 |a Artificial intelligence 
653 |a Success 
653 |a Python 
653 |a Trouble shooting 
653 |a Statistical analysis 
653 |a Clustering 
653 |a Safeguards 
653 |a Susceptibility 
653 |a Robustness 
653 |a Models 
653 |a Vulnerability 
653 |a Resilience 
653 |a Misinformation 
653 |a Urgency 
653 |a Unauthorized 
653 |a Language modeling 
700 1 |a Benjamin, Victoria 
700 1 |a Braca, Emily 
700 1 |a Carter, Israel 
700 1 |a Kanchwala, Hafsa 
700 1 |a Khojasteh, Nava 
700 1 |a Landow, Charly 
700 1 |a Luo, Yi 
700 1 |a Ma, Caroline 
700 1 |a Magarelli, Anna 
700 1 |a Mirin, Rachel 
700 1 |a Moyer, Avery 
700 1 |a Simpson, Kayla 
700 1 |a Skawinski, Amelia 
773 0 |t International Conference on Cyber Warfare and Security  |g (Mar 2025), p. 142 
786 0 |d ProQuest  |t Political Science Database 
856 4 1 |3 Citation/Abstract  |u https://www.proquest.com/docview/3202190576/abstract/embedded/7BTGNMKEMPT1V9Z2?source=fedsrch 
856 4 0 |3 Full Text  |u https://www.proquest.com/docview/3202190576/fulltext/embedded/7BTGNMKEMPT1V9Z2?source=fedsrch 
856 4 0 |3 Full Text - PDF  |u https://www.proquest.com/docview/3202190576/fulltextPDF/embedded/7BTGNMKEMPT1V9Z2?source=fedsrch