Overview
Learn about structured queries and defense mechanisms against prompt injection attacks in this 43-minute lecture from UC Berkeley's David Wagner at the Simons Institute. Explore how the lack of clear separation between instructions/prompts and user data creates security vulnerabilities in LLM-integrated applications. Discover a general approach to tackle prompt injection threats through explicit separation of prompt and data, while understanding how to modify standard instruction tuning to enhance model robustness. Gain valuable insights into alignment, trust, watermarking, and copyright issues surrounding Large Language Models through practical examples and implementation strategies.
Syllabus
Defense against prompt injection attacks
Taught by
Simons Institute