AI agents are quickly moving from simple chat interfaces to systems that can retrieve data, call APIs, access enterprise tools, and take actions on behalf of users.

In this session, we will build a simple AI agent, intentionally introduce common security weaknesses, and demonstrate how those weaknesses can be exploited using attack scenarios mapped to the OWASP GenAI Top 10.

The session will cover prompt injection, sensitive information disclosure, poisoned knowledge sources, excessive agency, unsafe tool usage, and improper handling of model outputs.

The goal is to highlight security patterns that apply across enterprise AI agents, copilots, RAG applications, and tool-enabled LLM systems.