Train the Team, Not the Analyst: An RL Gym for Securing AI SOC Agents