Security Patterns for Autonomous Agents: Lessons from Pentagi