Security Vulnerabilities in Large Language Models

This document is a presentation that outlines the security vulnerabilities associated with Large Language Models (LLMs). It begins by providing a high-level overview of LLMs from a security perspective, discussing how their design influences behavior and risk. The presentation emphasizes the importance of understanding potential misuse through a red teaming mindset, where security practices are employed to identify weaknesses in LLM systems. Key vulnerabilities are detailed, referencing the OWASP Top 10 for LLM applications, which includes issues such as prompt injection and sensitive information disclosure. The document discusses the development timeline of LLMs, explaining how they work and the significance of security testing in real deployments. It presents a layered approach to controlling LLMs, including data curation and runtime controls. Additionally, it covers the concept of red testing, which involves simulating attacks to discover vulnerabilities. The document concludes with assignments aimed at testing LLM vulnerabilities, including the use of tools like Ollama for security testing.