Security and privacy

We would rather not have your data than promise to protect it. Inference stays on your devices, connections go outbound, privileges are narrow and every call is on the record.

The hub and the nodes

Data stays in your network

  • Inference runs on your own nodes; the hub runs no models and stores no request bodies.
  • In the self-hosted edition nothing about your requests reaches us.
  • Remote image URLs are refused: the hub never fetches on your behalf (no SSRF).
  • With end-to-end sealing a client encrypts the job for the node that runs it; the hub relays ciphertext but still sees who, when, which model and the sizes.

Nodes only dial out

  • Nodes connect outbound over TLS (system trust store, or a pinned certificate for self-hosted hubs).
  • With "listen": false a node opens no port at all; it serves only jobs from its hub.
  • Nodes never route onward; forwarded requests carry a hop counter and loops answer 508.

Keys and least privilege

  • Virtual keys are stored as HMAC-SHA256 with a server-side pepper; a copied database cannot test guessed keys.
  • Client keys see only the models in their scope (others answer 404); node keys can only connect and report benchmarks.
  • Per-key rate limits and daily quotas; revoking a node key disconnects it immediately.

Everything audited

  • Every request: time, key name, client IP, method, path, model, node, status, latency, bytes, request id.
  • Never logged: keys, tokens or image data.
  • Proxy headers are believed only from configured proxies; CF-Connecting-IP only behind a Cloudflare hop.

Signed packages and releases

  • Model packages and recipes are signed with Ed25519 by a publisher key the hub and the node trust; the SHA-256 is checked before a package is served, and unknown manifest fields are rejected.
  • Every release ships a signed SHA256SUMS. The hub checks the signature on upload, the installer checks it again on the device and refuses downgrades.

Hardened hub

  • Distroless, non-root image (uid 65532) with no shell; read-only root and dropped capabilities in Compose.
  • Bounded bodies (10 MB default), sniffed content types, timeouts kept under the proxy’s.
  • Dashboard: random session in a Secure, HttpOnly, SameSite=Strict cookie, CSRF tokens, login lock-out, strict CSP, no-store.

Full details: the security model in the docs.

What this site does with your data

  • No analytics, no tracking or advertising cookies, no third-party scripts or fonts.
  • A strict Content Security Policy; scripts only from this origin. The live demo needs WebAssembly, so the policy allows wasm-unsafe-eval and nothing broader.
  • The waitlist form is the only thing that sends data, and only what you type, to our own hub. The live demo sends nothing.
  • Theme preference is kept in your browser’s local storage; nothing is sent to us.

Not done

  • No security certification (SOC 2, ISO 27001 or similar). None is claimed.
  • There is no hosted hub, so there is no relay we run for other people’s traffic.

Reporting a vulnerability

Please tell us via the waitlist form: start the “What would you run on it?” field with “Security”, describe the issue briefly and leave an email address we can reach you at; we will reply there for the details. Do not include anyone else’s data. This is a private beta run by one person, so replies may take a few days. See also security.txt.