The “high-speed insider” comparison really makes sense to me. Once an AI agent has tools, credentials, and enough freedom to act, traditional security assumptions can break down quickly. I think agent testing now needs to go beyond capability checks and include real security boundaries, monitoring, and containment from the start.