Comparative analysis of the vulnerability of large language models to prompt injections

Abstract

This article provides a comprehensive comparative analysis of the vulnerability of prominent Large Language Models (LLMs) to prompt injection attacks. It defines prompt injection as a critical cybersecurity exploit where malicious instructions, disguised as legitimate input, cause an LLM to deviate from its intended function. The report highlights that this vulnerability is ranked as the number one security risk by the Open Worldwide Application Security Project (OWASP) for LLM applications. The analysis is focused on the offensive methodologies and architectural weaknesses that enable these attacks.