GreyNoise doesn’t have much common need for detailed firmware analysis. If it’s happening on the internet, we already see it. However, when we do need to investigate vulnerabilities in embedded devices, things can get very complicated, very quickly if no information is publicly available. It can be fun and insightful to learn these skills in the rare case we need them.
In late October 2022, we became aware of CVE-2022-41140, a buffer overflow and remote code execution vulnerability in D-Link routers, which D-Link had been notified of on February 17th. Noting the months-long turnaround time, we decided this was a good chance to perform a learning and discovery exercise.
On March 13th, 2023 we became aware of CVE-2023-24762, a command injection vulnerability in D-Link DIR-867 devices. This recent CVE spurred us to share some of our internal documentation regarding a research spike into D-Link devices.
This blog aims to explain the process of gaining a foothold in firmware or a physical device for vulnerability research and achieving a debuggable interface. While existing Proof-Of-Concept code for (yet another) D-Link vulnerability CVE-2022-1262 is utilized within this document, as well as strong hints at suspect areas of code, don’t expect to find any new ready-to-fire exploits buried in the contents below.
What Vulnerability?
D-Link was notified of CVE-2022-41140, a buffer overflow vulnerability on February 17th, 2022. By November 15th, 2022, no additional information was available, which sparked an investigation into discovering available hints about the nature of the vulnerability. While this accurately speaks to the current state of public vulnerability tracking, we start off our investigation with a simple search on Google for the CVE and find two relevant links:
- https://www.zerodayinitiative.com/advisories/ZDI-CAN-13796/
- https://supportannouncement.us.dlink.com/announcement/publication.aspx?name=SAP10291
While the Zero Day Initiative lists the vulnerability as
(…) flaw exists within the lighttpd service, which listens on TCP port 80 by default. The issue results from the lack of proper validation of the length of user-supplied data prior to copying it to a fixed-length stack-based buffer.
the D-Link Technical Support page provides more detailed information
(…) a 3rd party security research team reported Buffer Overflow & RCE vulnerabilities in the Lighttpd software library utilized in DIR-867, DIR-878, and DIR-882/DIR-882-US router firmware.
A stack-based buffer overflow in the prog.cgi binary in D-Link DIR-867. A crafted HTTP request can cause the program to use strcat() to create a overly long string on a 512-byte stack buffer. Authentication is not required to exploit this vulnerability.
Additionally, the D-Link support page provides a table of the Affected Models
From this information, we can derive that the vulnerability is triggered by an HTTP request to TCP port 80, which will hit the lighttpd service and route to the prog.cgi binary resulting in an overflow on a 512-byte stack buffer.
We can also derive that the vulnerability can be patched/mitigated on some hardware models, but not others.
How to trigger the vulnerability?
The D-Link support pages provide links to download firmware images for the DIR-878, including base firmware versions like v1.30B08 as well as security advisement firmware versions like v1.30B08 Hotfix_04b.
Knowing that we can access the firmware images before/after the security patch for CVE-2022-41140, we will attempt the following steps:
Obtain copies of prog.cgi
We start by downloading a known vulnerable version of the firmware for a model that also offers a patched version. We download DIR-878_REVA_FIRMWARE_v1.30B08.zip and extract the firmware image DIR_878_FW1.30B08.bin.
We run the file command to quickly determine if it’s a commonly known file type. Unfortunately, this returns generic information.
Next, we use a more specialized tool, binwalk, which assists in searching binary images for embedded files and executable code. Again, this produces no results.
A handy feature of binwalk is the -E, --entropy command line flags, which allow you to measure the entropy or “randomness” of a file.
As an example, here is an entropy graph of 1024 bytes of Lorem ipsum:
And here is an entropy graph of DIR_878_FW1.30B08.bin
As you can see, the entropy of our firmware image is very high. Typically, this is indicative that a file is in a compressed archive format or is encrypted. Since neither file nor binwalk identified it as a compressed archive format, it’s reasonable to assume that it may be encrypted.
If you believe a file is encrypted, it’s always a good idea to take a peek at the bytes at the start of the file, just in case there’s an identifiable file header:
At the start of the file is a 4-byte sequence that maps to the ASCII characters “SHRS”.
A quick Google search for “SHRS firmware” turns up relevant results, indicating that we’re on the right track.
- https://github.com/0xricksanchez/dlink-decrypt/blob/master/dlink-dec.py
- https://0x00sec.org/t/breaking-the-d-link-dir3060-firmware-encryption-recon-part-1/21943
- https://0x00sec.org/t/breaking-the-d-link-dir3060-firmware-encryption-static-analysis-of-the-decryption-routine-part-2-1/22099
After a bit of reading, we can determine that D-Link does indeed encrypt some of their firmware, which is identifiable by the “SHRS” header. The blogs linked above go into depth on how they obtained a copy of the imgdecrypt binary and reverse engineer the binary to determine how to decrypt the firmware and produce the relevant python script.
Since we will be dealing with encryption again later in this blog, we won't go into depth on this specific layer of encryption. Our firmware can be decrypted with:
Taking our decrypted firmware and running it through binwalk again we can see that some file signatures are recognized.
Since file signatures were recognized, we can recursively extract them by using the -e, --extract, and -M, --matryoshka, command line flags.
This creates nested folders for each extracted layer of the file, ultimately resulting in a cpio-root folder containing the root filesystem for the firmware.
The desired prog.cgi file is located exactly where those familiar with *nix directory structures would expect it to be. However, for completeness, the file can be located by name using:
Now we have a copy of the entire root filesystem, including prog.cgi.
Repeating the same steps on the patched firmware sets us up for the next step.
Patch Diffing
In the previous step, we obtained an unpatched and patched copy of prog.cgi. We’ll rename them prog_old.cgi and prog_new.cgi, respectively, to help keep track.
BinDiff is a comparison tool for binary files, that assists vulnerability researchers and engineers to quickly find differences and similarities in disassembled code
For this blog, we’ll be using Binary Ninja with the BinDiff Viewer Plugin. There are roughly comparable free alternatives and plugins like Ghidra.
Following the relevant plugin steps to generate a bindiff, we open old/new and begin to look for functions that are very similar but not 1.00, indicating that a small change such as a patch may have been performed.
Uses of strcat()
Using our list of similar (but not exact duplicate!) functions, we work our way down the list, looking for uses of strcat() that have changed between old/new. In this example, the main function:
Old
New
Here we can see that the old binary used strcat() and the new binary has a different set of logic.
The strcat() function concatenates the destination string and the source string, and the result is stored in the destination string.
A quick check of the destination var_20c shows that its size is 0x200, or 512 bytes. For a sanity check, we can list all uses of strcat() throughout the binary.
There are 22 uses of strcat(). After reviewing them, none but the usage within main operate on a 512-byte buffer.
We now have a reasonable candidate for the location of the vulnerability.
Debugging with Emulation
Now that we have a reasonable candidate for a vulnerable code path, the next step is to start determining what conditions are required to actually reach the vulnerable code path. While wiser minds may be able to determine these conditions without needing a debugger, it’s always a safe bet to make getting a debugging interface a priority.
We want to run the necessary components and attach a debugging interface to a running program.
First, we need to determine the attributes of the file we would like to emulate. The file command we used earlier can be used to identify important information about the architecture the binary is meant to run on.
Using QEMU is an easy way to run binaries for other architectures, but with the same operating system as the current one. In this case, we want qemu-mipsel-static which is provided from the qemu-user-static package.
However, we need to know what to run.
There are init scripts that run when a system boots, and we can find the relevant one in /etc_ro/rcS:
It’s best to start at the top and work your way down and Google things where applicable.
- Filesystems are mounted
- /var/run folder is created if it doesn’t exist
- A script to create device (/dev) links is run
- The Message Of The Day (motd) is written to the device console
- A binary to manage reading/writing to non-volatile random-access memory (nvram) is started in the background
- A binary init_system is run with the start command
- A telnet daemon is started
/var/log folder is created if it doesn’t exist
Understanding the functionality of the init_system binary is elementary:
If init_system start is run, it checks for the presence of /var/run/nvramd.pid. If the pid file is not found, it enters a loop printing lighttpd: waiting for nvram_daemon. If the pid file is found, it branches into the following logic.
nvram is init then closed. sub_400e50 starts a number of .cgi binaries from /etc_ro/lighttpd/www/cgi-bin/, and finally the lighttpd web server is started with:
Using a combination of chroot and qemu-mipsel-static we can minimally and directly launch the lighttpd web service like so:
This results in an error of:
(network.c.747) SSL: Private key does not match the certificate public key, reason: error:02001002:system library:fopen:No such file or directory /var/private/lighttpd.pem
By simply commenting out the SSL related lines in /etc_ro/lighttpd/lighttpd.conf config file, we can just run the web service in HTTP mode exclusively and bypass the error.
Upon further review of the config, we can observe that the lighttpd web service is running in fastcgi mode and HTTP requests to the path /HNAP1/ are routed to be handled by prog.cgi.
If we navigate to our emulated system in a web browser, we can see that a page is loaded, and numerous UI assets load successfully, but the page is blank due to a malformed XML response from the /HNAP1 endpoint.
The root cause of the malformed XML response is due to the default values for nvram not being set. I spent a large amount of time trying to fix this by using LD_PRELOAD tricks and eventually ended up ordering a physical DIR-867 model (guaranteed vulnerable, no patch available) in frustration.
By the time the physical router was about to be delivered, I had a mostly working proxy for calls to functions from libnvram-0.9.28.so, at which point I remembered that the vulnerability was Pre-Authentication. I was trying to fix something that was part of the login flow which, I thought, was necessary.
After taking a bit of time to find a different endpoint to sanity check myself, I found that most of the other pre-auth functions of prog.cgi respond without issue. They are missing default values which would have been stored in nvram, but do not result in errors.
For our purposes, this is enough to work with and proceed forward. Getting a debugger attached by invoking prog.cgi in QEMU with the -g flag starts a GDB connection on port 1234.
Debugging With Physical Device
As stated earlier, I purchased a used physical model DIR-867 router, which is guaranteed to be vulnerable as no patches are available.
After opening the box, I began the setup process and set a device admin password of Password1 and set updates to “manual”.
After completing the setup steps, the router reboots.
Most importantly, I figured out how to reset the router using the button on the back and re-do the setup steps again to make sure nothing that I’d set so far has persisted across a reset.
Now that the router is set up in its most basic state, I do a quick scan for open ports.
Much to my chagrin, there is no 23/tcp open telnet result, despite the telnetd service appearing in the /etc_ro/rcS init scripts I’d found during emulation. I’ll need to find another way to get an interactive interface on the router to run a debugger.
At this point, physically opening the router up and trying to find a UART interface would probably be the quickest path to success. However, as I wasn’t in any particular rush, I decided to try to figure out how to just re-enable the telnet interface since I know from extracting the firmware that the telnetd binary already exists in the firmware.
Running a recursive grep on our extracted firmware shows that “telnet” shows up in many binary files, as well as what appears to be factory and default settings shipped with the device.
Note the telnetEnabled=0. This probably explains why telnet isn’t running. It also seems to indicate that it’s a setting.
While poking around earlier looking for command injection, I located the “System” menu, which allows exporting/importing settings. If we’re lucky, telnetEnabled is a hidden setting we can just flip on and re-import.
Clicking “Save Settings To Local Hard Drive” results in downloading a 5.9kB config.bin file.
We use binwalk to check for known file formats.
It looks like we have a SEAMA firmware header and the config is encrypted. Again, we turn to Google and search for the starting bytes of the file 0x5EA3A417 which returns a very useful C header file that defines the structure of a SEAMA file.
In the same folder on GitHub there’s a corresponding .c file for a command line tool to unpack a SEAMA file, but a quick review does not show any usage of OpenSSL. This likely means D-Link is doing some additional layer of encryption on top of SEAMA, and we’re better off doing some more static analysis on the firmware itself and using the GitHub repo for sanity checking ourselves.
Re-Opening prog.cgi in Binary Ninja and searching for usages of the string “config.bin” we see that it’s used in a single section of code at sub_42ad78.
Taking a closer look at sub_42ad78 we see the following flow graph:
At a high level:
- /tmp/config_2g and /tmp/config_5g are written to a manifest file /tmp/sysupgrade.conffiles
- The config files are put into a .tar.gz archive with tar czf "-" -T /tmp/sysupgrade.conffiles 2>/dev/null > /var/backup_tmp.tar.gz
- sub_42b2f4 reads model_name and hw_version from nvram and returns a string of “<model_name>_<hw_version>”
- The /var/backup_tmp.tar.gz file from step 2 is passed through a command mkconfig
- The resulting file is returned to be downloaded by the end-user
Taking a closer look at the call with mkconfig:
snprintf(&var_14c, 0x100, "mkconfig -a enca -m %s -i %s -o %s", &var_18c, "/var/backup_tmp.tar.gz", "/var/backup.tar.gz", 0x4f1530)
This results in &var_14c containing mkconfig -a enca -m DIR-867_A1 -i /var/backup_tmp.tar.gz -o /var/backup.tar.gz
Now that we know the command being run to generate the encrypted config.bin, we take a look at /bin/mkconfig to determine what those flags do. We can just run it in QEMU without any arguments to view the help message.
As the description states, it can encapsulate or de-encapsulate a config. However, it’s unclear where the suspected presence of encryption comes into play. A reasonable assumption from looking at the available flags indicates that the -m flag may be used in some sort of key derivation function. Remember that the -m flag is the model_name and hw_version. If the model and hardware version are used for a key derivation function, this would prevent someone from uploading a config from a different D-Link router model and potentially breaking their device.
We can confirm this by taking a peek at the enca function of mkconfig in Binary Ninja:
Indeed, we see the usage of openssl as well as a new, but fully expected binary seama.
In the first relevant part of the program flow, the -m flag (DIR-867_A1) is used in sub_400e30
Then the logic enters a loop that writes the MD5 hash as a hex string to &buffer
The OpenSSL command is as follows:
If we had preferred not to disassemble the function to figure out how the encryption key was generated, we could have simply added the -E "QEMU_STRACE=1" flag when running mkconfig and the resulting key would have shown in the strace output.
Command:
qemu-mipsel-static -E "QEMU_STRACE=1" /bin/sh -c "mkconfig -a de-enca -m DIR-867_A1 -i config.bin -o config.dec"
Strace output:
39 execve("/bin/sh",{"sh","-c","openssl enc -e -aes-256-ecb -k 81F9A6E40BDEC26DB67FE53A555D0E8E -in config.dec -out config.dec.enc >/dev/null 2>&1",NULL})
As expected, 81F9A6E40BDEC26DB67FE53A555D0E8E is the hex string representation of the MD5 hash of “DIR-867_A1”.
Knowing this is true, we can make a simple shell script to recreate this logic and patch in Telnet support:
- Use mkconfig to de-encapsulate (Unpack SEAMA firmware, Decrypt image)
- Extract the Gzip’d Tar archive
- Replace telnetEnabled=0 with telnetEnabled=1 in /tmp/config_2g
- Write /tmp/config_2g and /tmp/config_5g to a manifest
- Tar and Gzip the files in the manifest
- Use kconfig to encapsule (Encrypt image, Pack SEAMA)
The result being telnetpatched.bin, which should be a valid settings file for us to upload and enable telnet.
Indeed, another nmap scan shows the desired results of an open telnet port.
Unfortunately, when trying to connect, we are instantly prompted for authentication.
After a quick peek at the disassembly of prog.cgi, we can see that the password is set to the value we provided originally, Password1 + @twsz2018
A guess that the username is admin allows us to log in successfully with a password of Password1@twsz2018
While we have successfully logged in over telnet, we are dropped into a limited shell with only a select number of commands available to run. We cannot directly use this shell to load a gdb server and attach it to prog.cgi.
Here we will cheat a bit and leverage CVE-2022-1262, a command injection vulnerability in the protest binary that is available to us in the limited shell. Using the Proof-Of-Concept exploit included in this writeup from Tenable, we spawn another telnetd instance running on port 1337 and running as root.
From here, we can get a hint about the version of Linux headers the firmware was built with by running:
Finally, we can either cross-compile a mips32el uclibc GDB server against Linux headers 3.10.14+ ourselves by using something like crosstool-NG … or we can download a pre-built toolchain matching our criteria from https://toolchains.bootlin.com/
This allows us to transfer a gdb server to /tmp on the physical router and attach gdb to prog.cgi for remote debugging purposes.
Conclusion
Learning how to gain a debuggable interface with both emulated and real D-Link devices is a valuable skill for anyone interested in vulnerability research and network security. By using emulated devices, you can experiment and test in a safe environment before attempting changes to real hardware. The ability to debug real devices can help you identify and fix issues, as well as determine how firmware operates “in the real world”. While it may seem daunting at first, with the right tools and resources, anyone can learn these skills. With this knowledge, you can enhance your understanding of network security and device development, and apply these concepts to future projects, such as writing network signatures for malicious traffic.
GreyNoise tags for command injection CVE-2023-24762 and stack-based buffer overflow CVE-2022-41140 are live and available to all users for tracking related activity:
.png)
