We run a dedicated AI SRE for each instance with scoped creds for just their instance. OpenClaw by nature has security risks so we want to limit those as much as possible. We only provision integrations the user has explicitly configured.