Breaking and Hardening an AI Agent: A Full Red Team Lab on Indirect Prompt Injection, Data Exfiltration, and Defenses That Actually Hold