Multi-Chain Prompt Injection and Jailbreaking in LLM Applications - Security Testing and Defense Strategies

Learn about multi-chain prompt injection attacks targeting modern LLM application workflows in this technical presentation that explores vulnerabilities in systems using multiple LLM chains. Discover how traditional testing methods for jailbreak and prompt injection fall short when dealing with complex scenarios involving query rewriting, plugin interactions, and formatted outputs. Through detailed examples including a Workout Planner application, examine how attackers can exploit chain interactions to bypass security measures and propagate malicious prompts. Explore practical demonstrations comparing traditional jailbreaking tools like garak against multi-chain applications, analyze specific injection payload techniques, and understand potential mitigation strategies. Gain hands-on experience through the MyLLMDoc Challenge while accessing additional resources and documentation for implementing defensive measures in LLM applications.