Security you can inspect
Levelrail keeps your secrets encrypted, signs every request in with a session or a scoped token and never reaches into your servers over SSH after setup. The code is Apache 2.0 and open to read.
What is in place
Secrets are envelope encrypted
Each owner of secrets (an app, a backup target, email settings) gets its own random data key, and each value is encrypted under it with AES-256-GCM. Every data key is wrapped by one master key that is loaded into memory at start. Each value is bound to the slot it was written to, so ciphertext copied from one app into another fails to decrypt. Secrets are decrypted only at container create time and are masked on every read surface.
The master key stays out of your backups
By default the key lives in a file readable only by the service, or you supply it from your own secret store. It is never part of a control plane backup snapshot, and it can be wrapped with a HashiCorp Vault or OpenBao Transit key so that a copy of the data directory is useless on its own. Back it up separately.
Agents dial out over mutual TLS
Managed servers need no inbound ports and hold no SSH keys. A join token is exchanged for a client certificate, and with the control plane CA fingerprint pinned on the agent, an attacker in the network path cannot capture the token. After enrollment every connection is mutual TLS.
Scoped tokens and IAM policies
API tokens carry abilities, and a user can only mint a token holding abilities they hold themselves. Allow and Deny IAM policies scope to one app or database, and an explicit Deny always wins. Tokens are bound to the server that issued them, so a wrong URL or hijacked DNS name never receives one.
Sign-in hardening
Session cookies are HttpOnly and SameSite, and Secure over HTTPS. TOTP two-factor codes are single use, recovery codes are consumed on use, and passkeys are supported. The first admin needs a one-time setup token, so an exposed fresh install cannot be claimed by a stranger.
An audit log
Every mutating and sensitive request is recorded with actor, ability, method, path and caller surface (CLI, dashboard, MCP or API), and can be exported as CSV.
Release and install integrity
The installer fails closed on checksums. Release checksums are cosign-signed, and the installer can be told to require a valid signature.
Docker access is guarded
The control plane and each agent reach Docker through an allowlisting proxy that refuses privileged containers, host namespaces, unsafe capabilities and protected host mounts. It ships in enforce mode.
Outbound requests are fenced
Notification and webhook requests refuse to connect to loopback, private, link-local and other reserved addresses, checked on the address actually dialed and on every redirect.
What is not true yet
- Automatic HTTPS defaults to an internal, self-signed certificate until you enable public ACME (Let's Encrypt) in settings. Browsers warn until you do, and renewal has limited verification.
- No compliance certification is claimed, and the threat model lists known gaps.
- SAML single sign-on and SCIM are beta, off by default and not verified against a live identity provider.
- Anyone who can read your whole data directory and the master key can decrypt every secret. Keep them apart.
Levelrail states its limits openly. The threat model states what it defends against and where it does not yet, and the security center in the dashboard checks your own instance against the hardening steps.
Five things to do on a new install
- Serve the dashboard over HTTPS and set the dashboard URL.
- Give people the narrowest ability that works, and keep root for a few.
- Turn on two-factor sign-in for every account that holds root.
- Pin the control plane CA fingerprint on every agent.
- Back up the master key separately from the data directory.
Security questions
How does Levelrail store secrets?
Secrets are encrypted with per-owner data keys using AES-256-GCM, and the data keys are wrapped by a master key held only in the control plane process. The database holds ciphertext, and values are decrypted only at container create time.
Does it need SSH access to my servers?
No. Agents dial out to the control plane over mutual TLS and use the local Docker Engine API. Managed servers need no inbound ports for enrollment or day-to-day operation.
How do I report a vulnerability?
Use the repository's private security reporting on GitHub, or email support@levelrail.com. Fixed issues are listed on the security advisories page in the docs.