A CI runner that lives for months accumulates SSH keys, Docker images, and npm caches from other repos. That warmth is speed and a cross-job isolation failure. GitHub-hosted runners are ephemeral VMs per job (with documented software images). Self-hosted runners are often pets. The 2026 default for untrusted PR code is: ephemeral, one job, throw away the machine (or at least the container).

Buildkite's elastic CI stack and Actions runner scale-sets (Actions Runner Controller on Kubernetes, or cloud VM pools) are the same design: job arrives, mint a runner, job ends, terminate.

Why pets fail

Self-hosted actions-runner on a beefy box: cache is great; a malicious workflow (pull_request from a fork) can leave a systemd unit. Even trusted monorepos leak credentials into world-readable workspaces. Ephemeral runners still need OIDC to cloud (short-lived tokens) rather than long-lived access keys on disk.

yaml
permissions:
  id-token: write
  contents: read
# federate to AWS/GCP; no static keys on the runner

GitHub-hosted is simplest isolation and a cost/queue trade. Self-hosted ephemeral is for private networks, GPUs, or license servers. If you cannot terminate the VM, run the job in a disposable container and still assume kernel escape is in the threat model for hostile PRs — don't run fork PRs on privileged cluster nodes.

Caches without sticky disks

Actions cache and registry caches restore warmth without keeping the runner. Nix and container layer caches belong in object storage. Local Docker layer cache on a pet is why people refuse to go ephemeral; remote buildkit or registry cache closes that gap.

Label hygiene: jobs must request a runner label that only ephemeral pools have. A leftover self-hosted label on a laptop in a closet will pick up production deploys.

Fork PRs

pull_request_target with untrusted checkout is a classic hole. Ephemeral does not fix a workflow that checks out attacker code and then uses GITHUB_TOKEN with write. Combine isolation of compute with least-privilege tokens.

Read GitHub's hosted runner and autoscaling docs (and Buildkite's elastic stack if that is your queue). Then list every self-hosted runner by age. Anything older than a day that still accepts pull_request from forks is a standing incident.