Checklist
Checklist to vet hardware key remote access for OT security
Short answer
Accept a hardware key remote access product only on evidence: demand the tunnel protocol and cipher suite, prove the no-inbound-ports claim with a deny-all firewall test on a mirrored host, and keep the paired host in the industrial DMZ as the one door to the floor. Reject any product whose vendor cannot name its cryptography or whose revocation depends on the endpoint.
On this page
A hardware key that pairs to one plant host and promises remote access with no inbound ports and no VPN is still an encrypted tunnel between an endpoint you do not fully control and a machine inside your network. Accept it on evidence, not on the label. Demand the tunnel protocol and cipher suite in writing, prove the no-inbound-ports claim with a deny-all firewall test on a mirrored host, test endpoint isolation on a deliberately untrusted device, and put the paired host in the industrial DMZ — Level 3.5 in IEC 62443 segmentation — so the plant keeps one controlled door to the floor. The checklist below is the acceptance test: what to demand, what to measure, and what fails outright.
Judge the tunnel, not the label #
"Not a VPN" names a category, not a security property. A tunnel is a tunnel: its assurance comes from authentication strength, key management, cipher choice, replay protection, implementation quality, patch cadence and monitoring. A proprietary protocol receives less third-party scrutiny than a standard one, so it starts behind, not ahead; obscurity is not a control.
The endpoint caution that ICS-CERT attached to VPNs covers every transport equally: "VPNs are only as secure as the connected devices."
Demand in writing, before any pilot: the tunnel protocol, the cipher suite, the key-exchange method, the replay protection, and how keys are rotated and revoked. If the answer is "patented" or "government-grade" with no names attached, treat the product as an unverified encrypted tunnel and stop the evaluation there.
When the vendor also claims standards alignment, separate a certificate from a self-declared statement exactly as you would for a gateway: certificate versus compliance claim for gateways.
Put the paired host in the industrial DMZ #
IEC 62443 segmentation places remote-access components in the industrial DMZ — Level 3.5 — with SCADA and historian functions on the control LAN at Level 3 behind a firewall. A key-to-host product changes none of that. The paired host is the remote-access component, so it lives in the DMZ, and the firewall between DMZ and Level 3 carries explicit per-protocol rules — the same discipline as remote SCADA access through firewalls and proxies.
Treat the vendor's pairing server as part of your trust boundary, because the key-to-host mapping lives there. Record where it is hosted, who administers it, and what notification you receive when it is compromised.
Decision: pilot in the DMZ only. A design that puts the paired host on the control LAN, or lets the key reach controllers directly, is rejected at the paper stage. What absorbs a compromise of that one door is the defense behind it, which is why layered OT security controls protect a network path.
Prove the no-inbound-ports claim on a mirrored host #
The claim is testable in an afternoon, and the test record is the acceptance evidence. Build a mirror of the production paired host in a lab VLAN behind a firewall with one rule base: deny all inbound, permit outbound only to the vendor's documented endpoints, log everything, and capture on the host segment while a key connects.
Two mechanisms can deliver "no inbound ports": the host polls a broker for pending requests, or it holds one persistent outbound session. Both are outbound-first, and neither is inherently safer than a well-configured VPN — that is decided by the items in the first section.
Worked example. A 10-minute baseline with the key idle shows 20 outbound TCP sessions and zero accepted inbound connections: 10 min × 60 s = 600 s, at one poll per 30 s that is 600 ÷ 30 = 20 sessions. Scaled up, 24 h × 3,600 s = 86,400 s per day ÷ 30 s = 2,880 outbound sessions a day for the polling design. A persistent-session design instead shows one long-lived connection with keepalives every 15 s — 86,400 ÷ 15 = 5,760 keepalives a day — and firewall and IDS baselines must tell the two apart, because they produce different log volumes and different failure behavior.
Failure criteria: one accepted unsolicited inbound SYN fails the claim, and any destination the vendor did not document is a finding, not a feature. The mechanism behind the claim is ordinary how outbound-only tunnels traverse NAT — which is why it proves traversal hygiene, not security strength.
Test endpoint isolation on an untrusted device #
The strongest legitimate idea in this product class is endpoint isolation: the key bypasses the local operating system, loads its own video, keyboard/mouse and optional printer drivers in real time, and runs the session from the key itself. That reduces dependence on a possibly infected BYOD machine; it does not remove it.
Test it. Run the key on a deliberately contaminated laptop — a snapshot seeded with a known keylogger and USB malware — and attempt input injection, screen capture, driver hooking, and persistence after removal. Image the laptop afterwards and read the residue: USB device history, print spooler artifacts, memory captures, cached credentials. Drivers must be signed and must run at a named privilege level; a driver that demands administrative rights on an unmanaged machine fails.
Exercise the key lifecycle and attack the policy #
Lifecycle: enroll a key, pair it, revoke it, lose it, replace it, reissue it. Pass criteria: revocation takes effect within one polling interval — under 60 s at the 30 s example — the paired-host list cannot be edited without dual control, every change is attributable in a log the plant retains, and the hardware factor cannot silently degrade to password-only.
Then attack the policy: attempt file transfer, clipboard, printing, local drive mapping, screenshot capture, an unattended session, and a multi-hop from the paired host into the control network. Anything the plant has not explicitly approved gets blocked at the DMZ firewall, and that enforcement — not the key's good behavior — is the control. One safety rule regardless of product: no unattended session with write access to drives or motion; view-only, or an operator standing at the machine.
Acceptance checklist #
- Tunnel protocol, cipher suite, key-exchange method and replay protection named in writing.
- Patch process documented for key firmware and host agent, with a committed cadence.
- Pairing server treated as a trust boundary: hosting, administration and compromise-notification terms in the contract.
- Paired host deployed in the industrial DMZ (Level 3.5); DMZ-to-Level-3 firewall rules documented per protocol and port.
- Deny-all inbound test on a mirrored host passed: capture shows only outbound-initiated sessions to documented destinations.
- Endpoint isolation proven on a contaminated device: no input injection, screen capture, driver hooking or persistence; signed drivers; no administrative rights required.
- Key lifecycle exercised: revocation inside one polling interval, dual control on pairing changes, plant-retained logs, no silent fallback to password-only.
- Every bypass attempt (file transfer, clipboard, printing, drive mapping, screenshots, unattended sessions, multi-hop) blocked or explicitly approved and enforced at the firewall.
- Sessions attributable, time-limited, recorded where required, and terminable by the plant.
- Evidence pack kept with the change record: packet captures, firewall rule export, signed test report, key issuance log.
When to reject or redesign #
Reject outright: the vendor cannot name the tunnel protocol or cipher suite; revocation depends on the remote endpoint being clean; audit logs leave the enterprise before the plant can retain them; drivers require administrative rights on unmanaged machines.
Redesign and retest: any accepted inbound connection in the firewall test, any undocumented destination, any bypass that succeeded.
If a lost key, a vendor outage or a pairing-server compromise has no tested containment path, the product stays in the DMZ pilot with rollback to local-only operation predefined until one exists. Approve only a design that preserves the one door, keeps data and session policy in enterprise custody, proves endpoint isolation under test, and emits auditable events that operations and security can both review. Everything else is rejected.
Questions asked
Is a hardware key tunnel more secure than a VPN for plant remote access?
Not by label. Any encrypted transport between the key and the paired host is a tunnel, and its security comes from named cryptography, key management, implementation quality, patching and monitoring. Require the protocol, cipher suite and key-exchange method in writing as a condition of purchase; a vendor that will not name them has failed vetting, whatever the product is called.
How do I prove a no-inbound-ports claim before the plant is connected?
Mirror the paired host in a lab VLAN behind a firewall that denies all inbound, permits only documented outbound endpoints, and logs everything. A 30-second poll produces about 2,880 outbound sessions a day; a persistent session produces keepalives instead. One accepted unsolicited inbound connection, or any undocumented destination, fails the claim.
What should the contract cover beyond the technical tests?
A revocation service level tied to one polling interval, written notice when the pairing server is compromised, a log export format the plant can retain before anything leaves the enterprise, and a patch commitment covering key firmware and the host agent. The key-to-host mapping sits on vendor infrastructure, so these terms are part of the security design, not boilerplate.
When does the checklist have to be repeated?
At every change of key firmware, host agent version, pairing-server dependency, or firewall rule base on the access path. Keep the packet captures, firewall export and signed test report with the change record so the retest compares against a baseline instead of redoing everything with no reference.