Hardware signing

Vitrinode-HSM is a signing device built to hold one secret, Vitrinode's SSH user CA key, so that key never sits in plaintext on the server. It's a separate project being built alongside Vitrinode. This page covers what it's designed to do and how much of that actually works so far.

On this page: Why hardware· Status· What it is· What it won't do· How issuance is bounded· Threat model· Beyond v1

Why hardware

Vitrinode's SSH access (Phase 8, in development) issues short-lived OpenSSH user certificates from an internal CA. If the CA's private key is on the server, anyone who gets root on the server has the CA as well, and can issue certificates for any principal, with any lifetime, for as long as they like.

The HSM moves that one key onto a dedicated device. The device generates the key itself, never lets it out, and only issues OpenSSH user certificates that fit a fixed policy covering principals, validity and extensions. A compromised server can still ask for certificates within that policy, but it can't get the key or anything outside the policy.

Status

On real RP2350 hardware the device has signed certificates that authenticated against OpenSSH end to end, and it now generates and keeps its CA key. On-device provisioning, persistent storage, lifecycle handling and what happens if power is lost during provisioning have all been checked on a development board. Still to do: the rest of Phase 3 acceptance, trusted time, device audit, rate limiting, production sealing and server integration. All of the hardware testing so far has been on one RP2350 A2 development board. Verified on hardware means run and checked on a physical device. In development means implementation has started but acceptance is incomplete. Bring-up only means the hardware responds but nothing depends on it. Not started means specified but not implemented.

Vitrinode-HSM status as of 30 October 2026
Area Status Detail
Specification
VNHSM/1 protocol specification Frozen Architecture and protocol frozen at revision R4.
Byte-level wire contract & test vectors Frozen Phase 2 wire contract and golden binary vectors, frozen, plus a separate Phase 3 amendment and provisioning vectors for the persistent CA identity.
Firmware, verified on real RP2350 hardware
Secure / Non-secure isolation boundary Verified on hardware The TrustZone-M split, the typed Secure gateway and Secure-side fault handling.
SSH user-CA signing core Verified on hardware Builds and signs a real ssh-ed25519-cert-v01@openssh.com user certificate from typed parameters. A certificate from the device authenticated against stock OpenSSH in Phase 2 acceptance, and a principal outside policy is refused without using up a serial.
USB transport Verified on hardware Vendor-interface USB with automatic WinUSB binding on Windows.
VNHSM/1 end to end over USB Verified on hardware Report-identity, get-public-key and issue-certificate requests all work over the framed protocol, and malformed frames are rejected before they reach the Secure gateway (2 September 2026).
External real-time clock Bring-up only The RTC responds on the Secure bus and its oscillator runs. It isn't a trusted time source yet and nothing security-related depends on it.
Persistent CA identity, on a development board
Persistent CA storage Verified on hardware The keystore format is checked against an independent reference implementation. At boot the device works out the state of its keystore and loads a committed CA record across power cycles (Phase 3b, closed 3 October 2026).
On-device CA provisioning Verified on hardware The CA key is generated on the device and committed in stages, and you have to press the BOOTSEL button to prove you're there. A device that's already provisioned refuses to be provisioned again, and a verified factory reset wipes the CA (Phase 3c, closed 5 October 2026).
Lifecycle and power-loss recovery Verified on hardware Automated lifecycle checks on the board, including a corrupted keystore (Phase 3d.4). Power was also actually pulled at eight points during provisioning, and each time the device came back up in the expected state (Phase 3d.5, accepted 9 October 2026 over two runs).
Phase 3 acceptance In development Still to do: proving no secrets leak and the Non-secure probes on the Phase 3 image, resource limits, the full acceptance run, and a manual sign-off after a real power cycle.
Not started
Trusted time Not started A monotonic time floor and rollback protection. For now certificate start times come from the host, which isn't trusted, and that's for development only.
Production provisioning, sealing & secure boot Not started The one-way factory-to-locked lifecycle, using the one-time-programmable fuses, secure boot policy and firmware anti-rollback.
Tamper-evident audit log Not started The current firmware only has a per-boot counter in memory.
Rate limiting Not started Specified, not implemented.
Vitrinode server integration Not started The hsm signing mode, the client on the server side and the commissioning checks don't exist in Vitrinode yet.

Storage currently uses a development storage root, not production key protection. Its encryption and authentication checks prove the storage format and device binding work, but anyone with the firmware and a flash dump, or debug access, can get the CA key back. OTP-backed storage and locking down debug come later.

Vitrinode-HSM is a separate hardware project and it's still being built. It isn't part of the deployed system. Vitrinode's SSH certificates are currently signed by a development-only software CA, and the server has no HSM support yet. The device isn't sealed, hasn't been independently assessed or certified, and has only been tested on an RP2350 A2 development board. Vitrinode doesn't currently use a hardware security module.

What it is

A single-purpose device

Built on a Raspberry Pi Pico 2 (RP2350). It isn't a general keystore or a smart card. v1 holds exactly one key, the SSH user CA, and only does a few things: report its identity, return the CA public key, and issue an SSH user certificate within policy. The CA key is generated on the device during provisioning, which needs a physical button press, and is then kept across reboots.

TrustZone separation

The RP2350's Arm TrustZone-M keeps the code that touches the key apart from the code that talks to the host. The Secure world holds the key, re-checks every field of a request on its own copy, checks policy and does the Ed25519 signature. The Non-secure world only does USB and message framing, holds no secrets, and can only reach the Secure world through a thin, typed gateway. The idea is that a memory-safety bug in the USB or parsing code can't get to the key. Tests on the device confirm the key isn't in Non-secure memory or in any response.

Standard cryptography

Signing, hashing and random numbers use a pinned build of HACL, a well-known verified crypto library. The chip's hardware RNG goes through standard health tests and conditioning before it seeds an HMAC-DRBG.

Device-held clock

The production design has a battery-backed real-time clock inside the sealed enclosure, so the device has its own idea of the time that doesn't depend on the host. The prototype RTC has passed hardware bring-up, but trusted time isn't implemented yet. The plan is for certificate validity to be set by the device's own clock plus a monotonic floor it won't go back past, and not by whatever time the host claims it is.

Sealed enclosure

The production device is meant to ship sealed, with only USB and a status LED on the outside. No debug header, no boot override, and no way to re-key it once it's locked. It talks to the host using a framed binary protocol (VNHSM/1) over a USB vendor interface, one request at a time.

What it won't do

These are permanent. Later phases won't add any of them back.

  • No generic sign these bytes operation. The device builds the certificate itself from typed parameters and signs only that, never bytes the host has put together
  • No way to export the private key, with any command, in any lifecycle state
  • No PKCS#11 interface and no raw signer. Issuing a certificate is a specific request, not a general signing service
  • No firmware update over USB
  • No vendor master key or recovery backdoor

How issuance is bounded

The design assumes the host asking for certificates might be compromised, and limits what even a valid request can get. All of this is specified. Status has how much of it is implemented.

Policy envelope

Every request is checked against a fixed policy: which principals are allowed, the maximum validity, and which certificate extensions and critical options are permitted. Anything outside that is refused.

Device-held time

Certificate lifetimes come from the device's own clock and a monotonic floor, so a compromised host can't future-date certificates or build up a stock of long-lived ones by lying about the time.

Tamper-evident log

Every certificate issued is recorded in a hash-chained log on the device, and the server checks that against its own records. If certificates are issued during a compromise, it shows up afterwards.

Fails closed

There's no software fallback in the planned server integration. If the device is missing or isn't the one that was commissioned, issuing stops. That's a design requirement for now, as the server integration and commissioning checks aren't implemented.

Threat model

Listed by attacker. This is what the design is meant to handle, and Status has how much is built. The middle one is the reason the device exists.

Remote or network attacker

Fully in scope. The device is only reachable over local USB and only performs typed, policy-checked operations.

Root on the Vitrinode server

The reason this exists. The attacker can't get the key out. They can request certificates, but only within policy, rate-limited by design and with every one logged on the device, and a routine CA rotation cuts them off.

Casual physical access

In scope. Debug and boot override are meant to be fused off on the production device, and the external flash only holds wrapped data.

Skilled non-invasive attack

Partly covered. Fault-injection hardening and the choice of chip make it harder, but a single-fault bypass can't be ruled out completely.

Funded invasive lab attack

Out of scope. Assume de-capping and chip-level attacks will get the secrets off the device. The answer to that is operational: short certificate lifetimes, routine CA rotation, and noticing when a device has gone missing.

Beyond v1

v1 only holds one key, so its security boundary can be proven before the design gets any bigger. Possible next steps are signed device attestation and a recovery-device workflow, an SSH host CA, signed audit checkpoints, then a small multi-key store with a name-constrained device CA, and eventually issuing server TLS certificates. None of that is promised and none of it has dates.