Research insights and updates from the Fabraix team on AI agent security, adversarial testing, and RL safety.