Back Doors for Cross-Signed Drivers

Four undocumented registry values in one key vary the default validation of signatures on kernel-mode code such that Windows 10 may allow cross-signed drivers when it is otherwise documented as requiring Microsoft-signed drivers.

This may be welcome for running your own drivers on your own computers without having to send them to Microsoft. Or it may be an unwelcome exposure to software that would install drivers by surprise, including to let malware elevate from administrative access to kernel-mode execution.

Setting these values requires administrative access. Their action is subject to System Integrity policy, which provides the best defence.

Digital signatures on Windows kernel-mode drivers are usefully classified according to how directly the driver’s owner has to involve Microsoft in getting the driver signed: WHQL-signed; cross-signed; self-signed.

A WHQL-signed driver is signed with a certificate whose private key is kept by Microsoft so that only Microsoft can do the signing. WHQL means Windows Hardware Quality Labs. For many years the only way that Microsoft would sign a driver for an Independent Software Vendor (ISV) was if the driver was sent to Microsoft with a record of having passed an appropriate WHQL test suite. In those years, a WHQL signature gave some assurance of the driver’s quality.

For the many sorts of driver for which Microsoft had not yet devised tests for hardware compatibility, ISVs simply could not get WHQL signatures. Such drivers could instead be cross-signed by the ISV using both a Software Publisher Certificate (SPC) that is issued to the ISV by a third-party certification authority (CA) and a publicly available cross-certificate that Microsoft issues to the CA. In the particular way that Windows validates signatures on drivers, the signature on a cross-signed driver has a root certificate from Microsoft but it’s one that distinguishes the code verification as having been out-sourced. Microsoft’s involvement in cross-signing is only indirect, to vet CAs as having sufficiently high standards for authenticating that whoever they issue their certificates to is an identifiable (and hopefully responsible) software publisher. A cross-signature is some assurance that the driver, of whatever quality, is the work of a specific known entity.

Banning Cross-Signed Drivers

Of course, there are WHQL-signed drivers that don’t live up to the expected assurance of quality and there are cross-signed drivers that don’t live up to the expected assurance of authenticity. Even if no Microsoft-approved third-party CA ever issued certificates to frauds, some ISVs demonstrably did not guard the private keys in their SPCs carefully enough. Inevitably, kernel-mode drivers written as malware circulated with cross-signatures to look like the commercial products of legitimate (and even prominent) software vendors.

For Windows 10, Microsoft took the opportunity to draw a line and make it significantly more difficult for anyone outside Microsoft to sign drivers that Windows will load in anything like ordinary use. Microsoft’s first-known public statement of motivation for Driver Signing changes in Windows 10 was:

We’re making these changes to help make Windows more secure. These changes limit the risk of a driver publisher’s signing keys being lost or stolen and also ensures that driver publishers are strongly authenticated.

For most practical effect, all new drivers must be signed directly by Microsoft. The days of Windows users being exposed to cross-signed drivers that can have been signed by just about anyone, possibly for mischief, possibly without being traceable, are numbered. Though this goal surely could be met securely without going so far as requiring that drivers can load only if signed by Microsoft, there doesn’t look to be any reasonable dispute that the goal is worthy or even that cross-signed drivers would better be gone.

To support this new requirement of Microsoft signatures on all new drivers, the programme of signing by Microsoft only after passing WHQL tests is expanded so that a WHQL signature can alternatively be obtained after an attestation to Microsoft. Both for the ISV’s one-time registration with Microsoft and on each subsequent submission to Microsoft to get a driver signed, Microsoft requires a very strong verification of the ISV’s identity: not just any old certificate from the many CAs that Microsoft issues cross-certificates to but a more secure and expensive Extended Validation (EV) certificate from a handful of CAs that Microsoft yet approves for this new purpose. The bar is raised, if not directly for the quality of driver writing, then at least for the traceability of the writers.

Exceptions

Though cross-signed drivers are thus banished in general for Windows 10, there are exceptions. According to what Microsoft presumably intends as the definitive description of Driver Signing Policy, “Starting with Windows 10, version 1607, Windows will not load any new kernel-mode drivers which are not signed by the Dev Portal” but

Cross-signed drivers are still permitted if any of the following are true:

Let’s leave aside for now the mystery of Microsoft’s proclaiming this behaviour as new for the 1607 release when it’s very nearly a restoration of the original release’s behaviour which had been relaxed for the 1511 release (apparently without announcement). More or less continuous updates mean the world soon moves on, and mostly has by now. The confusion that Microsoft created for all this in the early releases can now be tipped into the dustbin of history, probably with no lessons learnt (either at Microsoft or in the community of device driver programmers that seemed unable even to want to free itself from the helplessness of depending on Microsoft for thin pickings). Let’s instead take Microsoft’s list of exceptions as definitive for the stated version and for the foreseeable future. How might these exceptions affect you?

New Commercial Driver

If you’re the writer of a new driver for commercial distribution, then the exceptions don’t much matter. Your only realistic choice is of what sort of WHQL signature, with or without WHQL testing, to get for your driver. If you tried instead to get your customers to configure their Windows for one of the exceptions as your price for their use of your cross-signed driver, then you would soon be out of business. And a good thing, too! Nobody should have to turn Secure Boot off for you, much as you would perhaps resent Microsoft telling you to turn it off for your final observations in private of your driver’s performance in real-world conditions before you send it to Microsoft for a signature.

Old Cross-Signed Driver

If you’re the writer of an old cross-signed driver that is already in commercial distribution and ought to work unchanged for Windows 10, then the exceptions have the doubtlessly welcome effect that your customers will not lose the use of your driver when they upgrade to Windows 10 and you’re not put to the otherwise unnecessary trouble of getting a WHQL signature. Say what you will about Microsoft, but their mostly very good record of attending to backwards compatibility is not appreciated nearly well enough, including to be recognised as a significant reason that Microsoft’s operating systems, rather than some others’, ever got to dominate the mass market (which backwards compatibilty anyway played a good part in building).

Own Driver

If you’re the writer of a driver for your own computers, whether for your private pleasure or because you write for internal use by an enterprise (so that “own computers” may mean hundreds, thousands or even hundreds of thousands), then you should want to know the exceptions because one or other of them may be acceptable to you in exchange for having to sign contracts with Microsoft and keep sending your intellectual property to Microsoft. I should not be surprised if more than a few government departments, security contractors, discreet financial institutions and secretive research laboratories have gone down this path, perhaps even with a nod and a wink from Microsoft. (See below about an NDA.)

Users

If you’re an ordinary user of computers—well, a security-conscious user—then you should want to know the exceptions because however much Microsoft’s raising of the bar might benefit the security of Windows users in general, it’s less than beneficial to your security in particular if your computer somehow becomes covered by one of the exceptions and you’re not immediately aware. Though it’s not, of itself, a security issue if Windows 10 loads a cross-signed driver, it arguably is if such drivers get loaded on your computer when you reasonably expected they would be prevented.

What’s reasonable to expect is hard to assess. Microsoft’s own descriptions of driver-signing requirements for Windows 10 have at times seemed impossibly confusing even to experts. For instance, Peter Viscarola, a “founder, President and Consulting Partner at OSR” with a long experience not just of device driver development but of close relations with Microsoft, seems more than a little frustrated in Yet another query about driver signing: “But to say this stuff is confusing is a considerable understatement.” This was in response to a wider problem, not just of knowing the driver-signing requirements for Windows 10 but of reconciling the varying requirements of multiple Windows versions in multiple configurations. But it was as late as November 2016, well over a year after arranging what might be thought a definitive Questions and Answers: Windows 10 Driver Signing with Microsoft’s James Murray, itself done as a welcome service to the OSR community after a long stream of observations such as “I agree it IS more confusing than it should be.” If even experts at writing drivers and educating driver writers have areas of confusion over these requirements, what hope does your typical system administrator, let alone an ordinary computer user, have of appreciating that a seemingly obvious interpretation hides all sorts of subtleties and caveats?

New or Upgraded

One subtlety is with how your Windows 10 might get to be treated as if it has been upgraded even though you installed it fresh. You might think the difference between new and upgraded is much like the difference between, say, Windows Home and Windows Professional, or even between Windows 10 and Windows Server 2016, and is fixed in stone or is at least tamper-proofed. And so it may be once Windows is fully up and running. But decisions about Code Integrity configuration have to be made not just before the kernel has initialised its support for tamper-proofing but before the kernel is even loaded. If only for now, these early decisions are made independently of the ordinary machinery for licensing. They thus depend on things that are not nearly as well settled or protected as you might think or hope.

The closest that Microsoft seems to have come to disclosing that the exception for upgraded systems is decided by something as simple and vulnerable as a registry setting is in a video presentation Driver Certification on Windows Client and Server by Scott Anderson on 27th May 2016. I transcribe from eleven minutes in:

And then finally, we are actually going to have a registry key. So, as I mentioned. And this registry key is, you know, intended just for testing, so we definitely don't want you setting that registry key as you install the driver. The registry key essentially mimics the same behaviour as if you have, say, an upgraded system. So, like I mentioned, if you have an upgraded system we continue to trust the cross-signed certificates and you'll be able to then test with your cross-signed certificates rather than needing to submit everything to us.

As far as I know, Microsoft never has published which “registry key” does this magic. According to Peter Viscarola, again, but this time in Signing during development and testing on Windows 10: “The registry key information is only available under NDA.”

I see no reason either to doubt him or to suppose that his own knowlege isn’t under the NDA. That disclosure by Microsoft under an NDA “means it’ll probably turn-up in lots of places online eventually” is at least plausible, if cynical, and a setting that’s named UpgradedSystem has indeed turned up on the Internet. I doubt, though, that it has turned up only because someone broke their NDA. As registry settings go, this one never can have been meant to stay secret for long. Its purpose is presumably that Microsoft sets it on all upgraded systems, and so its existence, though not immediately its meaning, will have turned up most readily just from comparing the registries of new and upgraded installations. (If there’s one thing the Internet is reliably good at, it’s to stimulate brute-force analysis by helping it circulate as some kind of research.)

It is anyway not safely inferred that UpgradedSystem is the “registry key” from the video. We should perhaps not parse anyone’s conference talk too literally, but see that what’s talked about “essentially mimics” an upgraded system, yet UpgradedSystem is less a mimic than the defining indicator for the Windows loader to know that Windows was upgraded. Also, the “registry key” is said to be something that Microsoft is “going to have”, yet UpgradedSystem was already old, having been introduced for the original release of Windows 10 (not that anyone outside Microsoft seems to have understood its significance so early).

Of course, such context is easier to read (or misread) into Anderson’s words if you already know that there are other registry values that he can have meant, including some that were not yet defined in public releases. Within the one registry key are as many as five values that govern whether a driver’s signature must be a WHQL signature to be valid. Most of these evidently aren’t set in the ordinary course of upgrading a Windows installation or even of using Windows. They therefore do not turn up in brute-force comparison of registries. For one, a Google search for the name in quotes gives just one match (today, 11th July 2018), and it’s just another dump from brute-force analysis but this time of plain-text strings in executables. These registry settings are as obscure as such things ever are.

Security

The obscurity of the registry values matters because of something else that Anderson says and for which he plainly intends the obvious direct meaning: “we definitely don’t want you setting that registry key as you install the driver.”

Well, no, of course we definitely don’t! The banishing of cross-signed drivers for fear that this relatively easy signing with third-party certificates has been hijacked by the unscrupulous would be completely undone if those unscrupulous writers can simply continue as before except to make their installer take the extra step of setting a magic registry value. That they won’t do this if they know how and it works is fanciful beyond words.

So, why, you should be asking, am I about to help them? Don’t I heed my own words? You’d certainly be right that I am about to document these registry values and I won’t disagree that my doing so may help the unscrupulous. If they didn’t already know the registry values, they will soon. If they can get administrative access to your computer, then they can set one of these (undocumented) registry values and then set more (documented) values that install their driver. For this very little trouble, they will gain a very good expectation that their driver will load when Windows 10 next restarts even if you have Secure Boot enabled and your Windows was never upgraded. If you ever then discover that you’ve given them kernel-mode access, you may be tempted to blame me for having revealed the means. So be it. I, of course, say to blame Microsoft for having provided the means and for having thought that you didn’t need to be warned—and while you’re at it, direct a little of that blame to those who signed the NDA and thus went along, happily or not, with Microsoft’s security by obscurity.

It’s well past time that Microsoft documented the applicable registry values. This is not to say that Microsoft ought to have documented the registry values from the start. Had I been writing this three years earlier, I’d like to think I’d have argued that what matters is not whether the registry values themselves are documented but that the need for and means of defence is documented. It would have sufficed had Microsoft documented that the exceptions are susceptible to being enabled by malware with administrative access and that Windows users may want to defend themselves by various mechanisms that Microsoft would document for easy use. But time has passed. The only ethical strategy this late from Microsoft is full openness—but do not expect you’ll ever see this!

What is a WHQL Signature?

Before I can list the registry values and describe their effects, we must sort out a question of terminology. What Microsoft’s programmers and technical writers sometimes mean by a WHQL signature when writing about driver signing for Windows 10 is not what programmers outside Microsoft have tended to think it means. As noted above, ISVs had been conditioned for a decade to think that getting a driver signed by Microsoft starts with getting it to pass the applicable WHQL test programme (if one exists). Given this background, it’s not the slightest bit surprising that when Microsoft writes some such phrase as “WHQL signature required” in a supposedly helpful table at yet another page on Driver Signing Policy, what ISVs read is “WHQL testing required”.

Now, I don’t mean to suggest that driver writers are frightened of testing, or to excuse them if they skimp on it, though I certainly do sympathise with their apprehension that WHQL testing may take them down a rabbit hole of overheads that far exceed the technical demands of designing and implementing a solid driver. Where the WHQL tests truly do test just an ISV’s driver, then the ISV can’t credibly be anything but glad that someone else has done the hard and unglamorous work of developing a test rig to use for free. There’s a good argument in such cases that the ISV would do better to run their driver through the WHQL tests even if not needed for driver signing (and even if the driver is for private use). Against this is that there have not always been—and there might not be for a good while yet—WHQL tests that truly do apply to all types of kernel-mode code that ISVs may think to write. Anyone who has even glimpsed the WHQL-testing experience that awaits the writer of a driver of a type that Microsoft has not explicitly supported for its WHQL tests will appreciate how alarming can be even the suggestion that drivers for Windows 10 must get WHQL signatures. If Microsoft didn’t mean that all new drivers for Windows 10 must pass WHQL tests, then Microsoft has been ham-fisted to talk of requiring WHQL signatures.

Still, “WHQL signature” plainly is Microsoft’s choice of term, in some contexts, for three types of signature:

This breadth, and particulaly its applicability to the attestation signature, shows in too many places with too much consistency to be anything but intended. For instance, public symbol files reveal Microsoft’s names for internal routines ImgpIsWhqlSignatureRequired in the boot manager and Windows loader and CipWhqlEnforcementEnabled in the kernel-mode Code Integrity module (CI), and if these routines determine that a WHQL signature is required or that WHQL enforcement is enabled, then an attestation signature is ordinarily satisfactory. This even carries through to the user-interface visibility of plain-text descriptions for the CodeIntegrity event log. Events 3082 to 3085 were added for the 1607 release to tell of failure to meet “WHQL requirements” and of “WHQL driver enforcement” being enabled or disabled, yet attestation-signed drivers ordinarily don’t fail the WHQL requirements and do load when WHQL driver enforcement is enabled.

Thus did I write above that the programme of signing by Microsoft only after passing WHQL tests is expanded so that a WHQL signature can alternatively be obtained after an attestation to Microsoft. As confusing as Microsoft’s terminology surely was for programmers who will entirely understandably have taken the term WHQL signature as suggesting WHQL testing, it’s too late to change now. We all have to stick with it. All I can hope is to clarify that it was Microsoft who made this murky.

In this view, and certainly for all this article, a signature that is obtained from Microsoft by attestation is a WHQL signature, but of a new type that comes with no WHQL testing and with slightly fewer legal obligations to Microsoft.

Taxonomy

As to what exactly Windows needs to see in a signature to recognise it as any sort of WHQL signature, I am somehow both astonished and unsurprised to find nothing public that I might simply refer to as definitive, or even as adequate, so that I can move straight on to listing those registry values.

WHQL Roots

It is, of course, well known that a WHQL signature must have one of relatively few well-known Microsoft certificates as the signature’s root. The boot manager, Windows loader and CI are each hard-coded with identifying details for different selections of Microsoft root certificates. For the following table, I reproduce the common names of the roots that Windows 10 requires for a WHQL signature:

Common Name Remarks
Microsoft Authenticode(tm) Root Authority old roots
Microsoft Root Authority
Microsoft Root Certificate Authority
Microsoft Root Certificate Authority 2010 new root
Microsoft Test Root Authority old root for WHQL Test Signature Program
Microsoft Testing Root Certificate Authority 2010 new root for WHQL Test Signature Program
Microsoft Development Root Certificate Authority 2014 (new) root for Insider Preview and flightsigning

In practice, the old roots won’t be seen in WHQL signatures that Microsoft gives to new drivers for execution on Windows 10. Of the new roots, Microsoft Root Certificate Authority 2010 is the one to expect in signatures for drivers that are intended for a formal Windows 10 release, rather than for quick fixes and previews. How the certificate chain gets to this root is irrelevant, but different types of signature will be created by signing with different certificates and an observed correlation will tend to be reliable for a time, such that a few examples can usefully be presented for concreteness:

Signature Type Certification Path (Root CA to End Entity)
built-in Microsoft Root Certificate Authority 2010 → Microsoft Windows Production PCA 2011 → Microsoft Windows
WHQL testing Microsoft Root Certificate Authority 2010 → Microsoft Windows Third Party Component CA 2012 → Microsoft Windows Hardware Compatibility Publisher
attestation Microsoft Root Certificate Authority 2010 → Microsoft Windows Third Party Component CA 2014 → Microsoft Windows Hardware Compatibility Publisher

Remember, though, that these are just correlations that are observed for real-world Windows installations at the time of writing. Microsoft is presumably free to change at will to whatever new certification paths take its fancy.

EKU Requirements

What does reliably distinguish the different signatures that are obtained from Microsoft in different ways but which all count as WHQL signatures is their Enhanced Key Usage (EKU). A signature is not a WHQL signature just for having one of the necessary Microsoft roots. Its end-entity certificate must have a distinguishing EKU—or even two of them. It seems tidier to tabulate the relevant EKUs before proceeding to what each means and in what conditions it applies:

EKU Symbolic Name Details tab in Certificate dialog
1.3.6.1.4.1.311.10.3.5 szOID_WHQL_CRYPTO Windows Hardware Driver Verification
1.3.6.1.4.1.311.10.3.5.1 szOID_ATTEST_WHQL_CRYPTO Windows Hardware Driver Attested Verification
1.3.6.1.4.1.311.10.3.6 szOID_NT5_CRYPTO Windows System Component Verification
1.3.6.1.4.1.311.10.3.20 szOID_WINDOWS_KITS_SIGNER Windows Kits Component
1.3.6.1.4.1.311.10.3.39 szOID_EV_WHQL_CRYPTO Windows Hardware Driver Extended Verification

The symbolic names are from WINCRYPT.H for each EKU as an Object ID (OID) in string form. Microsoft’s plain-text descriptions are from the MUI file for CRYPT32.DLL but are more easily experienced via the Digital Signatures tab in the Properties dialog for a file as shown in the File Explorer.

The EKU 1.3.6.1.4.1.311.10.3.5 is observed in Microsoft’s signatures on third-party drivers for Windows 10 whether the signature was obtained by WHQL testing or by attestation. This is as good a place as any to amplify the point that Microsoft plausibly has thought all along that an attestation signature is naturally a new type of WHQL signature. Attestation signatures are observed to have the additional EKU 1.3.6.1.4.1.311.10.3.5.1. Comments with the respective definitions in WINCRYPT.H are instructive:

// Signed by Microsoft through hardware certification (WHQL) 
// Signed by Microsoft after the developer attests it is valid (Attested WHQL)

Beware, though, that presence of the EKU with the appended “.1” in attestation signatures is just an observation. It seems to be true of Microsoft’s practice so far, but the boot manager, Windows loader and CI don’t care about this EKU. To them, any signature that has the EKU 1.3.6.1.4.1.311.10.3.5 is a WHQL signature (but read a few paragraphs further for an important caveat).

Anything that’s signed with a certificate that has the EKU 1.3.6.1.4.1.311.10.3.6 may as well be called Windows-signed. This term has support from the names of some internal routines as known from public symbol files, e.g., SbCheckPolicyIsWindowsSigned, which the boot manager, boot loader and CI all have for checking that a Secure Boot policy file is signed by Microsoft. Any signature that has this EKU counts as a WHQL signature.

In a few configurations, one more EKU, 1.3.6.1.4.1.311.10.3.20, suffices for recognising a signature as a WHQL signature. Examples of this EKU in signatures abound in the Windows Driver Kit (WDK) for Windows 10, but only for ARM and ARM64 builds. While I care only for general-purpose computers with x86 and x64 builds, I leave this EKU and the applicable configurations as a diversion that we don’t need for present purposes.

The EKU 1.3.6.1.4.1.311.10.3.39 does not suffice on its own for counting a signature as a WHQL signature. It can, however, be an extra requirement for 1.3.6.1.4.1.311.10.3.5 to suffice. This happens if a loaded System Integrity policy sets the “Required: EV Signers” option, as through the XML

<Rules>
  <Rule>
    <Option>Required:EV Signers</Option>
  </Rule>
</Rules>

When this policy option is active, presence of 1.3.6.1.4.1.311.10.3.5 in a signature makes the signature a WHQL signature only if 1.3.6.1.4.1.311.10.3.39 is present too. Or would a better way to put it be that a signature that has the first but not the second is still a WHQL signature but not a good enough WHQL signature for this stricter policy? The second EKU is observed in Microsoft’s signatures of third-party drivers for Windows 10 that seem to have got the signature through WHQL testing but not on any that look instead to have been signed by attestation. If this observation is indeed how Microsoft means to continue distinguishing these two types of WHQL signature, then there is an implication: when Windows operates under the “Required:EV Signers” policy, if the default validation of a signature on a kernel-mode driver requires a WHQL signature, then a signature that is obtained by attestation will be invalid and a driver that is signed only because of attestation will ordinarily not be allowed to load. I doubt that this is well appreciated by ISVs who get Microsoft to sign their drivers just by attestation.

Is a WHQL Signature Required?

Now that we know what a WHQL signature is, we can proceed to the registry values that influence whether a signature on a driver must be a WHQL signature if it’s to be valid for driver-loading.

Wouldn’t that be nice? But see that I say only influence, not determine. The registry values are not alone in the determination. Indeed, they mostly have the lowest precedence. I also don’t write simply “whether WHQL signatures are required” (though I will start to, for brevity). Most of the factors in the determination do apply generally, but there is one case that is evaluated per-signature. I also hedge around a signature’s being “valid for driver-loading” rather than commit to whether the driver is allowed to load. A driver with a signature that is valid for driver-loading may yet be rejected because of a System Integrity policy. A driver with no such signature may anyway get loaded (and not just because Windows is running under a kernel-mode debugger).

Indeed, while I can find in public no reliable overview of the algorithm for validating signatures on drivers, I also have to gloss over the circumstances in which Windows even cares for whether WHQL signatures are required. Notably, it does not come even close if validation is in the hands of a Secure Boot policy that specifies its own allowed signers. Even in default of that, which is usual, WHQL signatures don’t figure for validation if the signature is already invalidated for not having an accepted root. What counts as an accepted root is subject to some configurable options and varies with the type of driver and the circumstances in which it is being loaded. However important the details may be—and they are historically the most important part of the default validation algorithm—we can put them aside for present purposes except on one count. The effective root of signatures on cross-signed drivers is the Microsoft Code Verification Root. If in the circumstances, this happens not to be among the accepted roots, then signatures on cross-signed drivers are already invalid.

Even where a signature isn’t already invalid, the question of whether it must be a WHQL signature may simply be skipped. This happens for exotica such as HAL extensions and ELAM drivers (whose signatures will have long before been invalidated just as signatures unless they satisfy very particular EKU requirements). Starting in the 1703 release, it also happens if the accepted roots have been broadened, as for the testsigning boot option or because of Licensed Driver Signing, to allow custom roots. In effect, skipping the question means that a WHQL signature is not required. If a WHQL signature is not required, or if it is but the signature is indeed a WHQL signature, then the signature is valid for this stage and the driver is let through to a second stage which evaluates it against the loaded System Integrity policies.

The converse is our primary interest here and may as well be spelt out. If a WHQL signature is required but the signature is not a WHQL signature, then the signature is invalid and the driver will ordinarily not be allowed to load. This is where Windows 10 mostly rejects cross-signed drivers. To understand how it might instead allow them, and might be made to allow them, let’s take as given that validation of some signature has got as far as asking whether a WHQL signature is required. How is this particular question decided? As a rough guide, it’s decided by the following in decreasing order of precedence:

Factor Applicable Versions WHQL Signature Required?
Test Mode 1607 and higher no
applicable policy options all yes
upgrade in progress;
or Windows PE
1607 and higher no
one registry setting 1703 and higher yes
Secure Boot disabled;
or other registry settings;
or old signature
all no
default all yes, except in 1511

The details, inevitably, depend on whether the validation is done by the Windows loader or by CI and in which Windows 10 release. You should know that I, who have no help from Microsoft except for the formally released executables, documentation and public symbol files, am always behind the times, such that “and higher” presently extends only to the 1803 release. There is, inevitably, much that must still be worked out. What follows is as much as I yet know and have seen how to describe. If you want more, try getting it from Microsoft.

Test Mode

Starting with the 1607 release, WHQL signatures are explicitly not required if testsigning is enabled. I don’t yet see how this is not redundant in 1703 and higher (which, remember, avoid the question if the accepted roots have been widened to include custom roots). For 1607, however, it looks to have been a bug fix. The point to the testsigning option, perhaps even its reason for existence, is to allow the loading of drivers that are in development and are not yet signed for real-world use. The option is meant to boot Windows in a Test Mode that accepts drivers that have any signature at all, including most notably drivers that are self-signed with simple root certificates that a programmer might quickly make with the makecert tool. This is thwarted in the original and the 1511 releases if a policy option, as described shortly, would require WHQL signatures. Test Mode can have its intended effect only if it has a higher priority than the policy options, which it has in 1607 and higher.

Secure Boot Policy Option

In all releases, if the active Secure Boot policy specifies the “Required:WHQL” option, then of course WHQL signatures are required. Most computers, but perhaps not most tablets and phones, have only a default Secure Boot policy file embedded in the boot manager’s resources, and it does not specify this option.

As to what this option is properly called, I can only guess: Secure Boot policy files look to be completely undocumented. Microsoft surely prepares them from XML, plausibly from the very similar XML that acts as source code for compiling System Integrity policy files via PowerShell’s ConvertFrom-CIPolicy command. Though the binary format of a Secure Boot policy file is not that of a System Integrity policy file, each has a dword of bit flags for options and many of the bits are shared, as if the file formats were developed either in common or one from the other. Where the meaning of a bit is shared, as for this option (0x00000080), I take it as no bad guess that the name in the XML is shared too.

System Integrity Policy Options

If either the “Required:WHQL” or “Required:EV Signers” (0x00200000) option is set in any loaded System Integrity policy, then WHQL signatures are required. The difference, remember, is that the latter is stricter about what sort of WHQL signature is good enough. I can’t find a formal announcement by Microsoft that the “Required:EV Signers” option is intended as invalidating attestation signatures, yet drivers even from prominent vendors such as Lenovo are in circulation with WHQL signatures that were obtained by attestation and do not have the EKU that is required by this option.

By the way, you should be aware that System Integrity policies go by many names. I persist in calling them System Integrity policies because this looks to be what they were called by Microsoft’s programmers who wrote the code that interprets the policies. This is the name that survives not only in what the public symbol files tell us of relevant internal routines and variables, but also in the text of error messages, in the names that the files must have, as with SiPolicy.p7b, and in the names that header files from Microsoft’s programming kits assign to error codes, as with STATUS_SYSTEM_INTEGRITY_POLICY_NOT_SIGNED (0xC0E90004). In other contexts, Microsoft uses other names. Code Integrity policies, for instance, are preferred for the PowerShell commands that look to be the whole of Microsoft’s formal support for these policies. Elsewhere, Microsoft refers to Device Guard policies and even to Windows Defender Application Control (WDAC) policies. The last, despite being Microsoft’s latest preference, is conspicuously unhelpful for disguising its applicability to kernel-mode drivers.

Audit Mode

Whatever we’re to name them, these System Integrity policy options are not absolute. Among the many other options for System Integrity policies is “Enabled:Audit Mode” (0x00010000). This is intended to ease a policy’s preparation by logging what the policy would deny if not in audit mode. If “Required:WHQL” or “Required:EV Signers” would indeed be too restrictive for you by rejecting a driver that you actually do need, then learning about it by auditing the policy and inspecting the log very likely is easier than activating the policy and finding that Windows doesn’t start or that you lose some (peripheral) device without which Windows is difficult to use. Helpful though this crutch may be, it does mean that while “Enabled:Audit Mode” is set, including if it’s set to audit any of the many other things that System Integrity policies can affect, the “Required:WHQL” and “Required:EV Signers” options have greatly diminished effect.

For the original and 1511 releases, it would be as well to say the options are defeated. It just doesn’t matter that a “Required:WHQL” or “Required:EV Signers” option says WHQL signatures are required. The “Enabled:Audit Mode” countermands, with the end-effect that WHQL signatures are not required.

Starting with the 1607 release, however, audit mode actually does audit. The “Enabled:Audit Mode” option respects the “Required:WHQL” and “Required:EV Signers” options in that WHQL signatures are still required, but if a signature on a driver turns out not to be a WHQL signature, then the driver is granted an override for the audit. If all other things go as usual, the driver with its invalid signature is allowed to load but with an audit trail in the CodeIntegrity event log (specifically, as event 3082).

Upgrade in Progress or Windows PE

Also starting for the 1607 release, WHQL signatures are not required while Windows is being upgraded or is being run as the Windows Preinstallation Environment, e.g., as a recovery option. Whether an upgrade is in progress is determined from the registry:

Key: HKEY_LOCAL_MACHINE\System\Setup
Value: Upgrade
Type: REG_DWORD
Data: must be 1

Operation as Windows PE is indicated by the presence of the winpe boot option in the current BCD object. Though both do have the same effect as the documented exceptions, i.e., that WHQL signatures are not required even when Secure Boot is enabled, I list them here only for completeness, not to suggest that they are configurable ways around the usual requirement of WHQL signatures.

Registry Values in CI Policy

Of the exceptions that Microsoft documents for the 1607 release and higher, whether “Secure Boot is off in the BIOS” is not configurable in the registry, of course, but the other exceptions are—and have been from the very start.

Original Exceptions

All known releases of Windows 10 recognise a registry value as distinguishing that Windows was installed by upgrading:

Key: HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\CI\Policy
Value: UpgradedSystem
Type: REG_DWORD

Mere presence of the UpgradedSystem value means that WHQL signatures are not required.

Similarly old is a registry value that can bring forward the cut-off for whether a signature counts as old:

Key: HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\CI\Policy
Value: WhqlEnforcementDateTime
Type: REG_BINARY
Data: eight-byte time stamp

The data must be exactly eight bytes for interpretation as the usual Windows time stamp, i.e., in 100ns units since 1601. To be valid, the time stamp must be less than 0x01D0C991`830F4000, meaning midnight at the start of 29th July 2015, presumably to be understood as the release date for Windows 10.

Very plausibly, the WhqlEnforcementDateTime value only ever existed as an aid to testing in the lead-up to the first release, specifically to observe the different handling of what would count as old and new cross-signed drivers. The valid time stamp from the registry data, else 29th July 2015, acts as a cut-off for deciding whether any particular signature is required to be a WHQL signature. The time taken from the signature for this comparison is version-dependent. The original Windows 10 uses the signing time, if available, else the start time for the end-entity certificate’s period of validity. Only the latter matters for the 1607 release and higher. A signature that predates the cut-off is not required to be a WHQL signature.

I would not ordinarily spell out the converse, but there is a subtlety for the 1607 release when the validation is done by the Windows loader, as when a driver has 0 for its Start value. It is still true that if the driver’s signature does not predate the cut-off, then it must be a WHQL signature to be valid. What’s special is that if the signature turns out not to be a WHQL signature, then the driver is granted an exemption. If all other things go as usual, the driver with its invalid signature is allowed to load for apparently no greater price than a complaint in the CodeIntegrity event log (specifically, event 3083). My support for describing this handling as an exemption is that public symbols show that the corresponding event descriptor in CI is labelled WhqlEnforcementFailureExempted. That this case of a cross-signed driver being allowed to load is not among the exceptions that Microsoft documents for Driver Signing Policy, as quoted far above, is presumably because it was intended from the start to be short-lived. Indeed, where it is listed among the exceptions in the announcement of Driver Signing Changes in Windows 10, version 1607, Microsoft said outright that “Future versions of Windows will block boot drivers” and evidently meant it.

Relaxation (1511)

As noted above, WHQL enforcement for the original release of Windows 10 was relaxed for the 1511 release. Except if overridden by a Secure Boot policy or System Integrity policy, WHQL signatures are not required. In this release, the two original registry values continue to be read by both the Windows loader and CI, but their contents are immaterial and are ignored.

Restoration (1607)

For the restoration of WHQL enforcement in the 1607 release, Microsoft seems to have taken the opportunity not only to tweak the original exceptions and to enumerate them in public, but also to embellish the configurability. Starting with the 1607 release, Windows recognises two more values in the Policy key.

Key: HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\CI\Policy
Value: WhqlDeveloperTestMode
Type: REG_DWORD
Data: must be 1

When the WhqlDeveloperTestMode is 1, WHQL signatures are not required. This WhqlDeveloperTestMode value best fits a literal reading of Anderson’s description in Microsoft’s video. Of all the values in the key, only this one is named as if to exist “just for testing”. It does not itself indicate an upgraded system but “essentially mimics” UpgradedSystem. And it will have been, at the time, something that Microsoft was “going to have”. Perhaps now that the registry values are all disclosed in public without breaching Microsoft’s NDA, those who have signed the NDA are free to say.

Or perhaps not, since the information about which registry value is covered by the NDA is not information that is knowable, with certainty, except through the NDA. From outside Microsoft’s cone of silence, I can’t know that the NDA’s “registry key” is not instead the following, which looks to formalise multiple effects as generalised WHQL settings.

Key: HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\CI\Policy
Value: WhqlSettings
Type: REG_DWORD
Data: bit flags

A set 0x00000001 bit means that WHQL signatures are not required.

The 0x00000002 bit is a sort of opposite in the 1703 release and higher: if it is set, then WHQL signatures are required even if Secure Boot is off. See that it overrides all the other registry values (which are inconsequential unless Secure Boot is on). In the 1607 release, this role of a set 0x00000002 bit as an override of other registry values is all the bit has. It defeats the 0x00000001 bit and the UpgradedSystem value, but not WhqlDeveloperTestMode.

High bits in the WhqlSettings data consolidate other registry values in the Policy key:

Each can be set as its own registry value or as the corresponding bit in WhqlSettings to have exactly the same effect. This doesn’t mean, of course, that Microsoft intends that they ever be set as bits in WhqlSettings as a registry value, not that we know Microsoft’s intentions about setting anything in any of these registry values.

Not only do the WhqlSettings consolidate the other registry values and provide separately for disabling and enabling WHQL enforcement, but they survive to the user-interface visibility of the event log. The CodeIntegrity events 3084 and 3085, which record whether “WHQL driver enforcement” is enabled or disabled, qualify their report with two items of event data. The one that’s presented as Settings can have the 0x00000001 and 0x00000002 bits directly from the WhqlSettings and the 0x80000000 bit that corresponds to UpgradedSystem. (If WHQL driver enforcement is instead disabled because of WhqlDeveloperTestMode, then Settings shows as zero.)

As an aside, the other item for event data for CodeIntegrity events 3084 and 3085 is named Exemption. The 1607 release does not show it in the plain-text report of the event, but later releases do. The Exemption is ordinarily 1 whether WHQL driver enforcement is enabled or disabled. It is 0 only if WHQL driver enforcement is disabled because of a Secure Boot or System Integrity policy option.

Boot Drivers (1703)

The 1703 release recognises one more registry value in the Policy key:

Key: HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\CI\Policy
Value: BootUpgradedSystem
Type: REG_DWORD

As with UpgradedSystem, mere presence of the BootUpgradedSystem value is enough to mean that WHQL signatures are not required. The difference is that BootUpgradedSystem acts only for drivers that are loaded by the Windows loader. It is ignored by CI and thus has no effect on drivers that are loaded by the kernel (which by far the most are).

A first thought is that the BootUpgradedSystem value exists only as some sort of attempt at continuing something like the exemption that the 1607 release allows for boot drivers whose non-WHQL signature is invalid for not predating a cut-off. I’d hope, though, that there’s more to it. If BootUpgradedSystem is present on its own, then the 1703 and later releases will typically load cross-signed boot drivers but not load other drivers that are cross-signed. But the CodeIntegrity event log will show that “WHQL driver enforcement” is enabled and, unlike the behaviour in the 1607 release, there is no event for each boot driver that gets loaded despite having an invalid signature.

Defence

Although the registry values can be changed with administrative access, they are neither the first word nor the last in whether they allow cross-signed drivers. Simple options in one or another policy can require WHQL signatures such that the registry values in effect don’t matter. Even if registry values act so that a driver’s signature is not required to be a WHQL signature, the driver still must be evaluated against the full specification of what the System Integrity policies allow or deny.

The policy options thus provide a broad defence. The easiest way to prevent all cross-signed drivers from loading is with a System Integrity policy that has the “Required:WHQL” option (and does not have the “Enabled:Audit Mode” option) but otherwise allows everything. Presumably, Microsoft has a readily available example to download, along with clear directions for signing it for protection against the same malware or rude installers that might otherwise have set one of those registry values.

The obvious problem with this broad defence is that it’s no help to you at all if you actually do have an upgraded system and with it a dependence on some cross-signed driver or drivers. Of course, Microsoft might rather that everyone in this position should update those drivers—which likely would be beneficial—and then avail themselves of the broad defence. But if the ISV has not yet got an attestion signature or never will, whether for objecting to Microsoft’s procedures or for being out of business, then the broad defence is not available to you.

With the full power of System Integrity policies, you can specify how you are more fussy about signatures that Windows has otherwise approved, and you can accept or deny drivers as files identified by name, hash, or properties from its resources. For the question of cross-signed drivers that would otherwise be taken as valid because of those registry values, it will help you to remember that signatures on cross-signed drivers have the Microsoft Code Verification Root, which you can codify within a Signer tag as

<CertRoot Type="Wellknown" Value="08" />

You can then black-list or white-list. If you allow cross-signed drivers in general by allowing the Microsoft Code Verification Root, you can set up all sorts of other rules to deny specific signatures or drivers. Perhaps more usefully, just make it that the Microsoft Code Verification Root is neither allowed nor explicitly denied, and then set up other rules to allow just the cross-signed drivers that you actually do want.

Productive Use

White-listing is not just a finer defence against unwanted cross-signed drivers when you have an upgraded Windows that depends on particular cross-signed drivers. It is also a way that individuals and enterprises that write their own drivers can sign them for their own use without having to send them to Microsoft and without exposing their computers to cross-signed drivers from anyone else. Being able to use your own cross-signed drivers is of course a very poor cousin of getting Windows to accept drivers that you sign with your own certificates—but the rich cousin, integrated with Secure Boot for distinguishing your drivers from others’, is kept by Microsoft as something you need to be licensed for. Until this changes, the combination of setting a registry value for an exception to allow cross-signed drivers and installing a System Integrity policy to deny cross-signed drivers other than your own is as much control of driver-loading on your own computers as Microsoft permits you. If you dislike that, then you know who to petition.

Policy Preparation

Microsoft may yet publish formal documentation of System Integrity policies. Meanwhile, though Microsoft’s cook-book approach to describing what can be done with System Integrity policies is mostly directed at user-mode code integrity (UMCI) for controlling applications, enough is provided for getting by if you’re prepared to put in the time for trial-and-error or if you know by other means how the things work. I’d help more—and I may yet since documenting System Integrity policies is very firmly on my radar. But I also think that these back doors for cross-signed drivers are a mess of Microsoft’s making—indeed, just a small part of a big mess that Microsoft made over the whole subject of driver-signing requirements for Windows 10—and it’s properly Microsoft’s mess to sort out.