Overview
Learn about prompt injection vulnerabilities in LLM-powered agents through a detailed 19-minute technical video that examines the "Synthetic Recollections" research article. Explore key concepts including language models, SQL and prompt injection attacks, emergent abilities like chain-of-thought reasoning, and the ReAct (Reason+Act) framework. Follow along with practical demonstrations of ReAct agents in action and understand potential security risks through thought/action/observation injection examples. Gain valuable insights into building secure LLM agents using the OWASP Top Ten guidelines for Large Language Models. Download the accompanying mindmap and access comprehensive references including research papers, implementation examples, and security resources to deepen your understanding of LLM agent security.
Syllabus
- Introduction
- Prompt Injection Demo
- Table of Contents
- Language Models
- Injection Attacks SQL, Prompt
- Emergent Abilities Chain of Thought Reasoning, Reason+Act
- The ReAct Loop Agent, Executor, Tools
- ReAct Agent in Action
- Thought/Action/Observation Injection in ReAct Agents
- Building Secure LLM Agents OWASP Top Ten for LLMs
Taught by
Donato Capitella