The rapid adoption of Generative AI is transforming applications into autonomous systems capable of reasoning, accessing enterprise data, invoking external tools and executing actions on behalf of users. However, this evolution introduces a new attack surface that traditional application and network security controls were never designed to address.

Prompt injection, jailbreaks, system prompt leakage, sensitive data exposure, malicious MCP servers, excessive tool permissions and unauthorized agent interactions are emerging as critical security challenges for organizations deploying AI-powered applications.

In this session, we will explore the evolving threat landscape of agentic AI and examine why securing the model alone is insufficient. We will look at the security boundaries across the AI application stack, including user prompts, models, retrieval-augmented generation (RAG), APIs, MCP servers, external tools and enterprise infrastructure.

Through practical attack scenarios and architectural examples, we will demonstrate how attackers can manipulate AI applications, exploit excessive agent permissions and abuse tool integrations. We will also explore a defense-in-depth approach combining AI-aware security controls, identity and access management, network segmentation, runtime inspection, data protection and continuous monitoring.

Attendees will leave with a practical framework for identifying AI-specific risks, evaluating agentic application security and designing security architectures that enable organizations to adopt AI without compromising enterprise security.