LLM Security 101 - Risks, Attacks, and Mitigation Strategies

Dive into a comprehensive video lecture on LLM Security 101 presented by Trelis Research. Explore various security risks associated with Large Language Models, including jailbreak attacks, prompt injections, and malicious code. Learn how to detect and prevent these threats using tools like Llama Guard, custom classifiers, and Codeshield. Gain practical insights on server setup, user authentication, and response sanitization. Discover the impact of security measures on system performance and latency. Access valuable resources, including slides, scripts, and additional materials to enhance your understanding of LLM security best practices.

LLM Security Risks
Video Overview
Resources and Scripts
Installation and Server Setup
Jailbreak attacks to avoid Safety Guardrails
Detecting jailbreak attacks
Llama Guard and its prompt template
Llama Prompt Guard
Testing Jailbreak Detection
Testing for false positives with Llama Guard
Off-topic Requests
Prompt Injection Attacks Container escape, File access / deletion, DoS
1. Detecting Injection Attacks with a Custom Guard
Preventing Injection Attacks via User Authentication
37 Using Prepared Statements to avoid SQL Injection Attacks
Response Sanitisation to avoid Injection Attacks
Malicious Code Attacks
Building a custom classifier for malicious code
Using Codeshield to detect malicious code
Malicious Code Detection Performance
Effect of Guards/shields on Response Time / Latency
Final Tips
Resources