No home-made cryptography
All the cryptography comes from standard libraries or well-maintained, widely reviewed ones. Nothing is hand-rolled: no custom signing scheme, key derivation or authentication token.
These are the principles Vitrinode is built on. Anything that's still a design goal and not working code yet is labelled.
On this page: Principles· Limitations
All the cryptography comes from standard libraries or well-maintained, widely reviewed ones. Nothing is hand-rolled: no custom signing scheme, key derivation or authentication token.
A device's long-term private key is generated on the device during provisioning and never sent anywhere. The server signs certificates from the CSRs it's given and never sees a device's private key.
The server's TLS identity, the device CA and job signing each have their own keys. SSH certificates have their own CA key too. In the current development build that's a software key on the server, marked development-only. For production it's meant to live on a dedicated hardware device (Vitrinode-HSM), not the server.
A successful TLS handshake shows a certificate chains to Vitrinode's own CA. It doesn't show that the certificate or device hasn't been revoked since, so revocation is checked against the server's records on every agent request. Device trust state is separate again. It comes from attestation and decides whether a device can get privileged jobs.
Administrators sign in with WebAuthn/FIDO2 only. There's no password option in normal operation, and you can register more than one hardware key so losing one doesn't lock you out.
Sensitive actions need a recent security-key check, verified by the server. A client saying it's "recently authenticated" isn't trusted on its own. Step-up is currently required for device revocation, attestation re-baselining, adding a security key and creating the privileged systemd.restart job. SSH certificates go a step further. Each request gets its own check, tied to that request and only usable once, and synced passkeys are refused. A step-up check can only refresh a session belonging to the same user. Read-only jobs only need a valid session. Any privileged operation added later should need step-up as well.
A valid device certificate shows which enrolled device is talking. Whether that device is in a good state is a different question, and that's what attestation answers. Where there's a hardware TPM, attestation means a vTPM quote over measured-boot PCR state. Hosts without one get a weaker software-evidence check, and it's never shown as equivalent. The result is kept as the device's trust state, re-checked periodically, and stale or failed evidence downgrades it.
Privileged jobs are only created and dispatched for devices in the TRUSTED state. The executor on the host can't see attestation itself, so a job that was delivered before trust was lost can still run once within its few minutes of validity. Vitrinode records that in the audit log but can't stop it.
If someone got the SSH CA key they could give themselves access to every host, so it's the one secret that matters most. The plan is for it to live on a dedicated device (Vitrinode-HSM) that generates the key itself, never exports it, and only issues OpenSSH user certificates within a fixed policy. There's no raw signing operation and no PKCS#11 interface to misuse. On a development board the device issues real OpenSSH certificates and generates and keeps its CA key across reboots, though under development storage and not production key protection. Trusted time on the device isn't implemented yet, and it isn't connected to a server yet.
The agent that talks to the network runs unprivileged. Anything that needs root goes to a separate, tightly scoped executor over a Unix socket with permission checks and a fixed set of typed operations. If the agent is compromised, that alone doesn't give anyone arbitrary root.
The executor checks each job's signature itself, refuses jobs meant for another device, only reads its signing key and device identity from root-owned paths, and runs each job at most once, even across restarts. This has been tested and accepted on disposable hosts but isn't rolled out to enrolled hosts yet.
Remote operations are a typed, validated set, never free-form shell commands. Jobs are signed by the server, expire, carry a nonce so they can't be replayed, and are audited. If a break-glass shell is ever added it would be gated separately, need step-up and be heavily audited.
Registration, sign-in, sign-out, step-up, adding a credential, device revocation, attestation re-baselining, the job lifecycle (including refused jobs) and SSH certificate issuance and refusal all write insert-only audit records. Device enrolment and creating provisioning tokens aren't audited yet. The log is append-only at the application level. It isn't tamper-proof storage.
Job expiry and SSH certificate validity depend on clocks, so managed hosts and the server need synchronised time (NTP). The server doesn't trust host time. The executor refuses privileged jobs when the host reports its clock isn't synchronised, but that's a health check, not cryptographic proof of the time.
Separate keys, revocation checks on every request, signed typed jobs and privilege separation on the host are all implemented. They limit what a compromised component can do, but they don't make the server or managed hosts immune.
Vitrinode hasn't had a third-party penetration test or a formal compliance audit, and has no certifications. The hardware signing device (Vitrinode-HSM) is in development and isn't part of the deployed system. SSH certificates are currently signed by a development-only software CA.