Wednesday, December 23, 2015 At 12:02PM
Tis the season to be jolly… or so they say; but it is also the season to be wary and vigilant. At GDS we were recently discussing cold boot attacks against full disk encryption on Linux systems – it didn’t take us long to agree it was feasible, but just how hard would it be and how practical is it to execute an attack? After a little searching we didn’t come across any pre existing tools so there was only one way forward…
The Attack
Evil maid attacks can be used to target any operating system. For this research we focused our attention on Linux with LUKS full disk encryption.
Traditionally, when a Linux distribution is installed with full disk encryption there is a small partition that remains unencrypted to enable the decryption and bootstrapping onto the encrypted drive. This partition will be mounted at /boot and contains the kernel and the initial ramdisk (initrd). While it would also be possible to attack the bootloader or the kernel we decided to focus on the initrd image.
The initrd is a minimal set of tools for decrypting and mounting the root filesystem, along with any other tasks required to get other mounts ready (e.g /proc, /dev). Once the initrd has finished its job, it performs a pivot_root, during which it removes itself as the root filesystem and launches the main init script on the new (real) root filesystem. A lot of trust is placed on initrd, however, it is conceptually trivial to modify the initrd to run malicious code. The exact method for performing this might vary between different distributions, but it remains relatively simple.
The initrd is traditionally a gzip compressed cpio image. This is the case for the Debian based systems we tested, whereas RedHat based distributions (Fedora, RHEL, CentOS) now use dracut and consist of an uncompressed cpio image with a gzip compressed cpio image appended. Debian based initrds use ash shell scripts to perform the setup, whereas dracut based installations use systemd and its associated configuration method.
To perform our specific attack we chose to implement an LD_PRELOAD based bootkit due to the ease of development, but the same theory can be applied to injecting a malicious kernel module or other persistent binary. Our main aim in using the LD_PRELOAD method is to inject a shared object into the first binary that is executed on the freshly decrypted root filesystem. This is usually /sbin/init and will have the PID of 1. To do this, the easiest method is to just modify the init script to export this environment variable so that it is set when the pivot_root occurs. Because the filesystem changes it is also necessary to copy the shared object to the new filesystem at an appropriate time (post decryption). This can be achieved with the following two lines inserted inside the initrd’s init script just before the root switch is executed:
cp /hack.so /${rootmnt}/hack.so
export LD_PRELOAD=/hack.so
This works because the real root filesystem is decrypted and mounted under the temporary root mount prior to the pivot and the rootmnt variable is populated with the location of the mountpoint. However, before this takes place it is necessary to remount the target filesystem as read write as by default it is read-only. In our case we patched the init script where it parses the kernel commandline so that whatever argument was provided, the root is mounted readwrite. An alternative to this could be to just add mount -o remount,rw /${rootmnt} to our list of injected commands.
This injection point does not exist on dracut based initrds however, as the init executable is a binary rather than a shell script. This creates three issues of which we must bypass all to achieve injection into pid 1.
The first issue relates to copying our binary into the decrypted root filesystem. This is the easiest of the three to bypass. To achieve this we added two ExecPre directives to the systemd service file responsible for pivoting the filesystem root. This is roughly equivalent to inserting commands just before the script executes switch-root as before. The first command remounts the root filesystem as read-write (as it is read-only by default), and the second performs the copy.
The second issue is with LD_PRELOAD. As we are not modifying a shell script, and we cannot pass environment variables to this process (as it is called by the kernel), it becomes tricky to load our shared object. The easiest way to bypass this is to move the init binary to a different location, and insert our own shell script in its place, which ends by executing the original binary. Only 2 lines are required, the first to export LD_PRELOAD, and the second is to execute the original systemd binary. Note that this injection is into the initrd’s pid 1, not the eventual root filesystems pid 1. The reason we obtain this injection point is to help us with issue three.
The third issue is that before calling the switch-root command, the systemd environment clears out all environment variables using the clearenv() function. As this function is part of the standard library, it is possible for us to override this function so the injected process calls our function rather than the original. Because we do not care about actually clearing out the environment variables, we do not do a very thorough job. Our version of the clearenv() function will remove all environment variables, and inject our LD_PRELOAD variable into the environment. As the clearenv() function is only called once in this process (the time we want to break it anyway), this modification will not cause any side effects.
These three patches, when combined, will result in our shared object being copied to the encrypted root filesystem, and LD_PRELOAD being injected into the eventual pid 1 in our target filesystem, as in our Debian based example. Note that these are not security features, just that the systemd command pipeline is more complex than that of a flat shell script.
With this level of access, it is also possible to exfiltrate the user’s password provided for decrypting the root filesystem. Note, this does not require any weaknesses in the encryption, but subverts the password input processing.
In the case of Debian initrds, the password is piped between the executable that asks the user for their password and the executable that decrypts and mounts the root filesystem. We can just inject our own script in process pipeline to save the password as well as passing it on.
As for systemd, it uses more complex inter-process communication via Unix domain sockets. Rather than attack this part of the process, we chose to attack the actual mounting of the filesystem. Once again this is a library function, in a dynamically loaded library, called inside the systemd binary so it is possible to hook the function. The function that decrypts the drive is called crypt_activate_by_passphrase. This function takes, among other arguments, the password as a char array. By hooking this function we are able to access the the password. In this case, we want to wrap the original function so we use dlsym to access the real function and call it ourselves. However, before doing this we also save the password for retrieval later.
To ‘save’ the password, we simply append it to the shared object we later copy into the root filesystem. We chose this method because we already know that the file exists and will be copied, whereas we do not necessarily have any other mounts that will persist. This method also decreases the number of files we touch on disk. To access the password we then read the password from the end of our own file (located by reading the LD_PRELOAD variable) and set it as the value of the PASSWORD environment variable in the context of our reverse shell, so (in the case of meterpreter) the ‘getenv PASSWORD’ command will retrieve the password used to decrypt the drive. In our development, our proof of concept .so simply ran a python meterpreter reverse shell, as python was available by default on all of our targets.
Mitigations
There are a number of methods that can be employed to mitigate these issues. However, even using these mitigations there is no way to protect an attacker with physical access and enough time from reflashing the BIOS/UEFI.
The first mitigation would be to keep the bootloader, kernel and initrd on an external usb drive which is booted from in lieu of a /boot partition. Whilst possibly the most secure option, it is also the most awkward for the user as they must remember to always pull the drive when leaving the laptop unattended, including safely unmounting the /boot partition, if not already unmounted. It will also make updates more awkward as it is necessary to have the usb key present to update the initrd/kernel.
Another option would be to completely disable booting off external media. This should remove the possibility for automated attacks, but in some cases, such as with Debian initrds which contain a full shell, it may still be possible to manually mount and modify the initrd from inside the initrd itself. This could be done via an automated keyboard style device, which could achieve the same end without the requirement to boot off external media.
A slightly lesser option would be to enable the BIOS boot password. This should disable anyone from booting to anything without the BIOS password.
These last two options do not protect against the case in which the attacker has enough time to pull the drive and use their own laptop/device to mount the hard drive and modify it directly.
The final, and most secure option would be to extend SecureBoot into the initrd. As it stands, SecureBoot is able to verify from the bootloader to the kernel inclusive, but it stops short of the initrd. If this can be extended to also verify a signed initrd then it would be difficult to modify anything on the /boot partition without being detected. As above, there will still be a potential for attack if the attacker is able to flash the BIOS/UEFI to disable secureboot.
The best way to protect yourself from these kinds of attack would be to never allow your devices to fall into the hands of attackers. Our proof of concept attack was able to backdoor all targets in just over two minutes, but in a real world attack it may be possible to engineer a target specific payload that could take only a few seconds to run, if it only required copying a single file. If an attacker has physical access and just a little time we’ve demonstrated that if they come prepared this attack is more than feasible, so remember to keep your laptop safe this Christmas.
The code and more detail on our tool is available on our Github: https://github.com/GDSSecurity/EvilAbigail
Author: Rory McNamara
©Aon plc 2023
