A three-hour hands-on online workshop for security practitioners who want to build an AI agent that actually acts rather than advises. Participants start from a lightweight Python scaffold backed by a model hosted on Azure AI Foundry, connect tools, define the agent's objective, and point it at a deliberately vulnerable application inside an authorized lab. The workflow runs end to end: discover the target, exploit a vulnerability, then locate and transfer synthetic data to a controlled receiver. Along the way attendees inspect the raw HTTP requests, reproduce findings by hand, and map the attack chain to MITRE ATT&CK, with explicit attention to where autonomy works, where it fails, and when a human has to step in. Five stages cover getting started, reconnaissance, exploitation, collection and exfiltration, and a debrief, each opening with a live demo followed by guided lab work. Aimed at penetration testers, red teamers and security engineers; assumes basic pentesting, HTTP and authentication knowledge plus comfort with a terminal and light Python. Runs 9:00 AM to 12:00 PM Eastern, which is 6:00 to 9:00 AM Pacific.