AgentBreeder: Mitigating the AI Safety Impact of Multi-Agent Scaffolds via Self-Improvement