AI systems are increasingly embedded into critical business workflows, from copilots and chatbots to automated decision-making and agent-driven systems. However, these systems introduce a fundamentally new attack surface, where traditional security controls often fall short. Unlike deterministic software, AI systems can be influenced through crafted inputs, contextual manipulation, and adversarial interactions, leading to risks such as data leakage, policy bypass, and unintended actions.

This session takes a practical, attacker-first approach to AI safety and security. Through real-world scenarios and step-by-step attack walkthroughs, we will explore how techniques such as prompt injection, jailbreaks, and context manipulation can bypass safeguards in seemingly secure AI applications.

The session then focuses on defense. Attendees will learn a structured approach to red teaming AI systems, including identifying abuse cases, designing adversarial test scenarios, and evaluating system behavior across different risk dimensions. We will also cover defense-in-depth strategies for securing AI applications, including input and output controls, policy enforcement layers, secure system design, and architectural patterns that reduce risk and limit the blast radius of failures.

Attendees will leave with a practical framework to assess, test, and secure AI-driven systems, bridging the gap between traditional application security and the emerging challenges of AI safety.