Abstract

Through this workshop, participants will learn what continuous red teaming means and how it can protect AI agents from security vulnerabilities like prompt injection, data leakages and business alignment failures like hallucinations, inappropriate denials, etc. We'll present how enterprises can automate security evaluation, detect vulnerabilities before they become incidents, and ensure continuous protection of AI agents. Through hands-on exercises aligned with Giskard Hub, attendees will explore a variety of testing techniques including context-specific prompts, adaptive probing, and multi-turn interactions.