Skip to content

Security

DasCondor edited this page Nov 15, 2026 · 2 revisions

Security

Documentation snapshot from docs/SECURITY.md at commit 20d7c81. Check the current source for updates.

KubeFactory · All docs · 2.0.0-beta.2 pre-release · Known limitations

Reporting a problem

Please report security problems privately through GitHub's "Report a vulnerability" button on the repository's Security tab, not in a public issue. Include what you did, what happened, and the version (shown in the Esc menu).

What KubeFactory can and cannot do

KubeFactory needs no server and has no credentials of its own (server mode, SERVER.md, is optional). It talks to Kubernetes through its companion program (kf-agent, started by the game, using your kubeconfig) and to Docker through the docker already on your machine, as you. It can do exactly what your own account can do, and nothing more.

  • Your kubeconfig is read, never copied out. The companion reads it to connect. Its contents are not written to any log, not shown in full in the game, and not sent anywhere except, as credentials, to the cluster it names. Pasted kubeconfig text is stored in your user folder only.

  • Sign-in helpers need your approval. A kubeconfig can name a program to run for credentials. KubeFactory will not run it until you approve that exact command for that profile, and never runs anything else an imported file names.

  • Storage inspection is read-only and confined. It lists and reads files inside a pod's mounted volumes using fixed commands in the pod's own container. It refuses paths that leave the volume, never reads Secret or service-account volumes, never creates a pod and never adds a mount. It needs the pods/exec permission; without it you get "not permitted". File previews are shown as text or hex and are never run or rendered as HTML.

  • LiveView is a separate, disposable browser. It uses the Edge or Chrome on your machine, headless, in a profile created for that display and deleted when it closes. It can only load the host you opened; everything else is refused. It has no access to your own browser's cookies or saved passwords. Port-forwards it uses listen on 127.0.0.1 only.

  • Your cluster permissions are the security boundary. The in-game tool tiers, protected namespaces and typed confirmations are guard rails against mistakes. Anyone with kubectl can bypass the game, so anything that must not happen has to be prevented with Kubernetes RBAC. See ADMIN.md.

  • Presence and the shared action feed are self-reported. A user allowed into the kubefactory namespace could post under another name. The authoritative record of who did what is your API server's audit log.

  • The browser demo connects to nothing. It only runs the built-in simulation.

What appears on screen

KubeFactory is often shown to other people (screenshots, streams, a shared screen), so it keeps details of your computer off the screen:

  • Any path under your own user folder is shown as ~ (for example ~/.kube/config), in panels, notes and error messages alike.
  • On a cluster you are shown by the name the cluster gives you, which everyone on that cluster can already see. On the local Docker engine, where there is no such identity, you are shown as local user rather than by your Windows account name. In the simulated worlds you are DasCondor.
  • Your account name is still written to the audit log on this computer, and sent to an audit webhook only if you configure one: that is what they are for.
  • tools\bench.ps1 leaves the graphics card model out of the published benchmark table unless you pass -ShowGpu.

Grafana and Tautulli tokens

  • A token is typed once, handed to the companion, and sealed with the Windows Data Protection API for your Windows account. The game keeps no copy, no operation returns it, and it is not in providers.json in readable form. On systems without that protection a token is refused rather than stored in the clear.
  • Links the game opens (a dashboard, a panel) carry no credential. The page runs in LiveView's disposable browser and asks you to sign in yourself.
  • Only the provider's own address is contacted; redirects are not followed. Plain http is accepted for this computer, or for another one only after you tick the box that says the token would travel unprotected.
  • Tautulli: one read-only request (get_activity), from which only totals are taken. No user names, addresses, titles or history are read out, and there is no way to send any other command. The key goes in a header; the key-in-address mode for Tautulli older than 2.18 is off unless you turn it on, because that server may then log the key.
  • Changing a provider's address drops its saved token: a token belongs to the address it was entered for.

Foreman's AI providers

  • Off by default: Foreman works from rules alone, and the browser build has no way to reach a provider.
  • API keys are handled like provider tokens: typed once, sealed by the companion for your Windows account, never returned, logged or shown. An error a provider sends back has the key removed before the game sees it.
  • A provider whose address is not this computer is a cloud provider. Nothing is sent to it until you set Cloud: ALLOWED for that provider. Switching providers never resends a question, and an outage never falls over to another provider.
  • What is sent: the question and a summary the game builds (names, states, counts and the Krew's findings), with passwords, tokens, keys, URL credentials, private keys, long encoded strings and your home folder removed. Show what is sent displays it. No logs, files, kubeconfig, socket or credential is given to the model.
  • The model's reply is untrusted. It is told that the context is data, not instructions, but the protection does not depend on that: a proposed action must name a tool from a fixed list and a resource that exists, becomes the game's own typed operation, and runs only when you press its button, through the tiers, read-only mode and confirmation. There is no shell tool.
  • Bounded: 48 KiB per question, 8192 answer tokens, 120 s, 2 MiB reply, one question at a time per provider, no retries, redirects not followed.

How secrets are handled

  • Authentication is whatever your kubeconfig and Docker setup already use. KubeFactory keeps a list of the kubeconfig files you added (their paths) in connections.json in your user folder. If you paste kubeconfig text instead of pointing at a file, that text is saved in the same folder, as plain text like any kubeconfig: protect it the same way. Removing the profile in the wizard deletes it.
  • Generated database passwords go straight into a Kubernetes Secret (via a short-lived temporary file, never the command line). On a plain Docker engine they are written to a per-project .env file in your user folder.
  • API keys, basic-auth users and webhook URLs you put in portainer.json, log-providers.json or policy.json are passed to curl through a short-lived config file rather than command-line arguments. Those files are plain text: protect them like any other file holding a key, and do not commit filled-in copies.
  • An API key is only ever sent over https, and TLS certificates are verified unless you set "insecure_tls": true in portainer.json.
  • The audit log records actions, not secrets.

App lists are untrusted input

App lists come from the internet. Before anything from a list is built:

  • every template is validated (image name, paths, setting names, restart policy, port protocols) and refused if anything is malformed;
  • nothing is ever passed through a shell, so template text cannot run commands on your machine;
  • host-path mounts (for example the Docker socket) are left out unless you tick an explicit box for that template.

A template still decides which container image runs. Only build from lists and images you trust, as you would with docker run.

Downloads

The setup script and the build workflows verify the Godot engine and export templates against pinned SHA-512 checksums. Releases ship with a SHA-256 checksum per package and one SHA256SUMS, and each package has a build provenance attestation signed through Sigstore by the release workflow (gh attestation verify <package> -R DasCondor/kubefactory checks it came from this repository's workflow at a given commit). No package is signed with a Windows or Apple developer identity: Windows SmartScreen warns on first run, and the macOS app is only ad-hoc signed (not notarized), so Gatekeeper blocks it until the quarantine flag is cleared. The companion's Go dependencies are pinned by agent/go.sum.

Clone this wiki locally