TroutTrout
Back to Blog
CMMCArchitectureCUIOT SecurityRemote Access

The Last Hop: Carrying CUI From a Cloud Enclave to the Machine That Uses It

Trout Team9 min read

The short version

A cloud CUI enclave solves storage, processing and collaboration. It does not solve the moment a controlled drawing has to become a physical part.

That last hop, enclave to printer, enclave to CNC, enclave to test bench, is almost always bridged by a person downloading a file to a laptop. At that instant the CUI leaves the boundary you assessed and lands on a device with a different control set, and the gap between the System Security Plan and the observable behaviour of the shop floor opens up.

The architecture below closes it: a FIPS 140-3 validated tunnel from the enclave into an on-site protocol-aware proxy that sits in front of the equipment, so the controlled path runs to the machine and back, with nothing installed on the machine.

Why the last hop is hard

HOW THE LAST HOP IS USUALLY BRIDGEDCLOUD ENCLAVECUI AT RESTDOWNLOADLAPTOPCUI AT REST, UNPLANNEDPRINTER / CNCNO IDENTITYCONTROLLED BOUNDARY ENDS HEREBROKERED, BOUNDARY CONTINUOUSCLOUD ENCLAVECUI AT RESTFIPS 140-3 VPNACCESS GATE APPLIANCEVM 1 ALMALINUX, VPNVM 2 PROXY, IDENTITYSESSION RECORDEDJOBRESULTSPRINTER / CNCUNCHANGEDCONTROLLED BOUNDARY RUNS TO THE ASSET AND BACK
The same journey twice. Above, the boundary ends at a laptop nobody assessed. Below, it runs to the asset and back.

Three properties of production equipment make it resistant to everything an enclave assumes:

  • It cannot authenticate. Machine tool controllers, network printers, coordinate measuring machines and bench instruments mostly predate the idea that a device should prove who it is. No agent to install, often no certificate store, no path to MFA. The enclave's identity model has nowhere to land.
  • It cannot be changed. The controller is validated as part of a process. Touching its configuration can invalidate a qualification, and in regulated manufacturing the cost of revalidation dwarfs the cost of the control you were trying to add.
  • It is reached by protocols that assume a trusted network. LPD and IPP for print, FTP and SMB and vendor-specific transfer for machine tools, raw sockets for much instrumentation. None were designed for a hostile network and none carry identity.
  • It is physically accessible to people who are not in your directory. Contractors, calibration technicians and OEM service engineers stand in front of these machines. Whatever control you add has to survive someone with hands on the panel and a USB port.

So the two usual answers both fail, and they fail differently. Downloading to a laptop moves CUI onto an endpoint nobody scoped. Extending the VPN to the plant converts a narrow requirement, move this file to that machine, into broad network reachability that is hard to bound and worse to document.

How the three approaches compare

Where CUI comes to rest
Download to laptop
An endpoint outside the assessed boundary
Extend VPN to plant
Wherever the operator saves it
Brokered through a proxy
Nowhere; it is never written down en route
What the far end can reach
Download to laptop
Whatever the user can reach
Extend VPN to plant
The subnet, and onward
Brokered through a proxy
One asset, one protocol
Identity checked against
Download to laptop
The laptop's own login
Extend VPN to plant
The tunnel, once, at connect time
Brokered through a proxy
Every session, before the asset is touched
Works on a device that cannot authenticate
Download to laptop
Only because a person carries the file
Extend VPN to plant
No, the device still trusts the network
Brokered through a proxy
Yes, enforcement sits in front of it
Evidence produced
Download to laptop
None
Extend VPN to plant
Connection logs, not actions
Brokered through a proxy
Per-session record of what was done
Return path for derivative CUI
Download to laptop
Undefined, usually a shared drive
Extend VPN to plant
Undefined
Brokered through a proxy
Same path, reversed
How it reads in an SSP
Download to laptop
An exception you would rather not write
Extend VPN to plant
A large boundary that is hard to justify
Brokered through a proxy
One documented path per asset

The architecture

Four elements, in order along the path.

1. The enclave holds the CUI. Nothing changes about how the enclave is built or assessed. It remains the system of record for controlled data and the place users work.

2. A FIPS 140-3 validated tunnel carries it. A site-to-site tunnel runs from the enclave to the on-site appliance, and it is worth being precise about which component holds the validation. On the appliance, an AlmaLinux VM builds and maintains the VPN, using the Kernel Cryptography Module for AlmaLinux 9 (CloudLinux Inc., TuxCare division), FIPS 140-3 Overall Level 1, under CMVP certificate #4750, validated 2 August 2024 and updated 7 August 2025. Access Gate runs as a separate VM on the same appliance and does the proxying described below.

That split matters for the paperwork. The requirement in play is SC.L2-3.13.11, employ FIPS-validated cryptography to protect CUI in transit, and it attaches to the cryptography carrying the data, not to the proxy deciding who may send it. Keeping the two functions in separate VMs means you can point an assessor at the component the control is actually about.

State it in that shape, because that is the shape an assessor can check: name the module, name the certificate, and claim that the product uses a validated module. NIST validates modules and issues certificate numbers, it does not certify products, so "FIPS certified" is a phrase that will cost you credibility in the room.

3. A protocol-aware proxy terminates the session on site. This is the load-bearing element. The proxy does not route the enclave onto the plant network, it terminates the session and opens a separate, constrained one to a single device. It understands the protocol in use, so authorization is expressible as "this identity may send a print job to this printer" rather than "this subnet may reach that subnet on port 9100". The session is recorded at the protocol level.

4. The device receives the job unchanged. No agent, no firmware change, no reconfiguration. From the machine's point of view a normal client on the local network sent it a normal job.

This is the shape of the Trout Software and CUI Vault integration. CUI Vault, from NtelSec, runs the enclave as a managed Microsoft GCC High environment, Access Gate runs the appliance in the plant or the lab, and the two connect natively, so neither side needs per-site engineering to open the path.

The consequence worth stating plainly: CUI never rests on an intermediate workstation. There is no download step to describe, restrict or monitor, because there is no download.

The return path, which is where the gap usually is

Most designs of this shape are drawn left to right and stop at the machine. That leaves the more awkward half undefined.

Equipment produces output. Inspection results, CMM measurement files, as-built records, process logs. When that output derives from controlled technical data it is commonly derivative CUI, and it is created outside the enclave, on a device with no controls, often landing on a shared folder or a USB stick because nobody specified anything better.

The same path runs in reverse. The proxy collects output from the device and returns it into the enclave, under the same identity model and the same recording. The record of what was built ends up stored under the same controls as the drawing it was built from.

If you take one architectural point from this post, take that one. A CUI journey specified only in the outbound direction is a boundary with a hole in it, and it is the hole nobody notices until an assessor asks where the inspection reports live.

Enduring exceptions, used correctly

There is a temptation to claim the CNC now "meets" the access control requirements. It does not, and writing that down is how an SSP becomes something you cannot defend in an assessment.

The honest construction is the opposite. The asset cannot meet the practice as written and will not be able to. That is what an enduring exception is for: a documented, persistent inability of a specific asset to satisfy a requirement, recorded with the compensating control that manages the resulting risk. The proxy is the compensating control. It is what enforces identity, least privilege and recording on behalf of an asset that can do none of those things itself.

Two things make this hold up under scrutiny:

  • The exception names the asset and the requirement, not a general category. "Legacy equipment" is not an exception, it is an excuse.
  • The compensating control is evidenced, not asserted. Because the proxy sits in the path, it produces the evidence as a by-product: who connected, to which device, over which protocol, when, and a replayable record of the session. That is materially stronger than a policy statement about how the machine is used.

One SSP instead of two

The organizational payoff is documentation, and it is bigger than it sounds.

Without this integration, a defense manufacturer typically maintains a cloud enclave with one boundary, one control set and one assessment, and a shop floor that is either out of scope by assertion or in scope and poorly described. The connection between them is a person with a laptop, which appears in no diagram.

With it, the boundary is continuous and the diagram is honest: enclave, validated tunnel, proxy, asset, and the return path. It is one System Security Plan describing one path, with the exceptions named where they genuinely exist and the compensating controls evidenced where they apply.

Where this fits in CMMC

Level 2 is where most of this conversation happens, and the architecture is aimed squarely at the access control and audit families that an assessor will spend time on.

Level 3 layers a subset of NIST SP 800-172 enhanced requirements on top, for the highest-priority programs. Those requirements push toward stronger isolation of CUI, tighter control over what may reach it, and better evidence of both. An architecture in which controlled data never lands on an intermediate endpoint, every transfer is bound to an identity, and every session is recorded, is pointed in that direction. Scope the specifics with your assessor rather than assuming a clean mapping; the useful claim is about the direction of travel, not a checklist.

A note on timing, since it changes the urgency and not the requirement: CMMC Phase II was suspended in July 2026, so no third-party certification deadline is currently forcing the calendar. The underlying NIST SP 800-171 and DFARS obligations did not go anywhere. The controls still apply, and the last hop is still where most programs are quietly out of conformance with their own documentation.

Related reading

FAQ

Frequently Asked Questions

Where does a cloud CUI enclave actually stop?
At the human. An enclave is excellent at storing, processing and sharing CUI among people on managed endpoints. It has no answer for the moment CUI has to become physical: a drawing printed, a toolpath loaded onto a machine tool, a test profile pushed to a bench instrument. In most programs that last hop is made by downloading the file to a laptop, which is where the controlled boundary ends in practice even when the SSP says otherwise.
Why a proxy instead of extending the VPN to the plant network?
A VPN terminates on the network and hands the far end a route. That turns a CUI transfer into general reachability of the plant, which is the opposite of what you want and is very hard to describe in an SSP. A protocol-aware proxy terminates the session instead, so the enclave reaches one named device over one named protocol for one job. The unit of access becomes the machine rather than the subnet.
What does FIPS 140-3 validated mean, and is it the same as FIPS certified?
They are not the same, and the second is not a thing. NIST does not certify products. Its Cryptographic Module Validation Program validates cryptographic modules against FIPS 140-3 and issues a certificate number for that module. The correct claim is that a specific module is FIPS 140-3 validated under a given CMVP certificate. In Access Gate that module is the Kernel Cryptography Module for AlmaLinux 9 from CloudLinux Inc., TuxCare division, FIPS 140-3 Overall Level 1, CMVP certificate #4750. Assessors notice the difference, so say validated, name the module, and be able to produce the certificate number.
How can a twenty-year-old CNC meet CUI access control requirements?
It cannot, and pretending otherwise is how SSPs become fiction. The controller has no identity, no MFA and often no useful logging, and it will not get them. The workable approach is to stop trying to make the asset meet the practice and instead enforce the practice in front of it: identity and authorization checked at the proxy, the session recorded there, the asset unchanged. That is a compensating control, and where the asset genuinely cannot meet a requirement as written it is documented as an enduring exception with the compensating control named against it.
Does this apply to CMMC Level 3?
It is relevant to it. Level 3 layers a subset of NIST SP 800-172 enhanced requirements on top of Level 2 for the highest-priority programs, and those requirements push toward stronger isolation of CUI, tighter control over who and what can reach it, and better evidence of both. An architecture where CUI never lands on an intermediate workstation and every transfer to a device is identity-bound and recorded is aimed at that direction of travel. Scope it with your assessor rather than assuming a control maps cleanly.
Is inspection data CUI?
Frequently yes, as derivative CUI. Output generated from controlled inputs, CMM results, as-built records, inspection reports and some machine logs, commonly inherits the marking of the technical data it came from. This matters architecturally because that output is created on the equipment, outside the enclave. A CUI path built only in the outbound direction leaves the return trip undefined, which is usually the real gap.