Truecrypt 7 Derived Code/Windows: Drive Letter Symbolic Link Creation EoP
Class: Local Elevation of Privilege
Tested on: VeraCrypt 1.13 x86 on Windows 10
The Windows driver used by projects derived from Truecrypt 7 (verified in Veracrypt and CipherShed) are vulnerable to a local elevation of privilege attack by abusing the drive letter symbolic link creation facilities to remap the main system drive. With the system drive remapped it’s trivial to get a new process running under the local system account.
Any user on the system can connect to the Truecrypt device object and mount a new encrypted volume. As part of this process the driver will try and create the requested drive letter by calling IoCreateSymbolicLink. To prevent redefining an existing drive letter a call is made to IsDriveLetterAvailable which attempts to open the link “\DosDevices\X:” for reading, returning FALSE if it was successfully opened. The specific code in src\Driver\Ntdriver.c is:
if (NT_SUCCESS (ZwOpenSymbolicLinkObject (&handle, GENERIC_READ, &objectAttributes)))
The bug which allows you to bypass this is due to the use of the NT_SUCCESS macro. This means that any error opening the symbolic link will cause the drive letter to be assumed to not exist. If we can cause the open to fail in any way then we can bypass this check and mount the volume over an existing drive letter. This is possible because with terminal services support the \DosDevices path points to a special fake path \?? which first maps to a per-user writable location (under \Sessions\0\DosDevices) before falling back to \GLOBAL??. When the kernel creates a new object under \?? is creates it in the per-user location instead so there’s no conflict with a drive symbolic link in \GLOBAL??. So how to bypass the check? The simplest trick is to just create any other type of object with that name, such as an object directory. This will cause ZwOpenSymbolicLink to fail with STATUS_OBJECT_TYPE_MISMATCH passing the check.
This in itself would only cause problems for the current user if it wasn’t for the fact that there exists a way of replacing the current processes device map directory using the NtSetInformationProcess system call. You can set any object directory to this which allows you DIRECTORY_TRAVERSE privilege, which is pretty much anything. In particular we can set the \GLOBAL?? directory itself. So to exploit this and remap the C: drive to the truecrypt volume we do the following:
1) Set the current process’s device map to a new object directory. Create a new object called C: inside the device map directory.
2) Mount a volume (not using the mount manager) and request the C: drive mapping. The IsDriveLetterAvailable will return TRUE.
3) Wait for the driver to open the volume and at that point delete the fake C: object (if we don’t do this then the creation will fail). While this looks like a race condition (which you can win pretty easily through brute force) you can use things like OPLOCKs to give 100% reliability.
4) The mount will complete writing a new C: symbolic link to the device map.
5) Set the \GLOBAL?? directory as the new process device map directory.
6) Unmount the volume, this calls IoDeleteSymbolicLink with \DosDevices\C: which actually ends up deleting \GLOBAL??\C:
7) Remount the volume as the C: drive again (you’ll obviously need to not use C: when referring to the volume location). The user now has complete control over the contents of C:.
Fixing the Issue:
While technically IsDriveLetterAvailable is at fault I don’t believe fixing it would completely remove the issue. However changing IsDriveLetterAvailable to only return FALSE if STATUS_OBJECT_NAME_NOT_FOUND is returned from the ZwOpenSymbolicLink object would make it a lot harder to bypass the check. Also I don’t know if specifying the use of the mount volume driver would affect this.
The correct fix would be to decide where the symbolic link is supposed to be written to and specify it explicitly. As in if you want to ensure it gets written to the current user’s drive mapping then specify the per-user directory at \Sessions\0\DosDevices\X-Y where X-Y is the authentication ID got from the SeQueryAuthenticationIdToken API. Or if it’s supposed to be in the global drive names then specify \GLOBAL??. Note this probably won’t work on pre-fast user switching versions of XP or Windows 2000 (assuming you’re still willing to support those platforms). Also I’d recommend if going the per-user route then only use the primary token (using PsReferencePrimaryToken) to determine the authentication ID as that avoids any mistakes with impersonation tokens. There’s no reason to believe that this would cause compat issues as I wouldn’t expect the normal user tool to use impersonation to map the drive for another user.
Note this wasn’t reported in the iSec Partners security review so it’s not an missed fix.
Proof of Concept:
I’ve provided a PoC, you’ll need to build it with VS2015. It will change an arbitrary global drive letter to a VeraCrypt volume. Note it only works on VeraCrypt but it might be possible to trivially change to work on any other truecrypt derived products. You MUST build an executable to match the OS bitness otherwise it will not work. To test the PoC use the following steps.
1. Create a veracrypt volume using the normal GUI, the PoC doesn’t do this for you. Don’t mount the volume.
2. Execute the PoC, passing the drive letter you want to replace, the path to the volume file and the password for the file. e.g. MountVeracryptVolume C: c:\path\to\volume.hc password.
3. If all worked as expected eventually the PoC should print Done. At this point the drive letter you specified has been replaced with the truecrypt volume. As long as you have a command prompt open you should be able to see that the C: drive is now pointing at the encrypted volume. You can hit enter to exit the program and unmount the volume, however if you’ve replaced the system drive such as C: this will likely cause the OS to become unusable pretty quickly.
It shouldn’t be possible to mount the volume over a global drive.
The global drive specified has been replaced with a link to the encrypted volume.
Proof of Concept: