27 Sep

Why I’m not using x64dbg

x64dbg is (probably) the most user-friendly x64 debugger right now. It’s pretty, it’s open-source and it usually works. But I find it very hard to switch from WinDbg to x64dbg for several reasons. Some of them are purely emotional (don’t worry, I’m not going to bore you to death explaining those) but most of them are technical and related to the way x64dbg is being developed.

So, here goes slightly exaggerated but still serious list of my grievances. smile

Insane system requirements

Both DNSpy and x64dbg suffer from this disease. They love to use the “latest and greatest” of technologies, meaning Visual Studio 2017, .NET 4.6 and what not. That’s perfectly fine when you’re writing normal software. But debugger is not a normal software.

If I have a customer with a software crashing on his production servers, I can’t tell him “You need to install Windows 7 SP2 and 3 different VS redistributables and reboot your machine twice just for me to run my debugger”. No, I really can’t.

Debugger must run on any and all systems out-of-the box. Olly does that. WinDbg does that. And it wouldn’t be hard to link x64bdg with static VS runtime libs and target WinXP while using all the modern goodies. But for some reason it’s not done that way.

Updated 30-Sep-2016: Mr. eXoDia let me know that now x64dbg is distributed together with the necessary runtime DLLs. We can remove this grievance off of my list. Hooray! smile

Uncertain direction and feature bloat

Antoine de Saint Exupery said:

Perfection is finally attained not when there is no longer anything to add, but when there is no longer anything to take away

These are really wise words and Olly is designed that way. It does all the basic stuff and has stable SDK that enables plugin authors to implement all the extras.

On the contrary, Mr. eXoDia is adding features left and right and the direction of x64dbg development looks more like this:
BrownianMovement

For example, why does a debugger need 3 (yes, three!) different assembler engines?
assemble

Want another example? Let’s just look at the latest weekly digest.. How about this:

… change to the info box. Basically the pointer values in the instruction were not resolved (so if the instruction contained qword ptr ds:[rsp+30] it would not show the value of rsp+30). Personally this is quite useless

Yes, Mr. eXoDia, you’re right. It is useless for everyone but few people.

And how about:

The commands plugload and plugunload have been added. This is useful for plugin developers who want to test plugins without having to restart x64dbg all the time.

How many people in the entire world will actually benefit from that? 5? 10?

So, why add such bloat? Once you add something, that something must be maintained. And it’s very hard to remove stuff later, as it might break something else. So, please don’t..

Broken features

When I am on a job and need to debug something, last thing I want to spend my time on, is fighting with debugger bugs. And my customers certainly don’t want to pay me for doing that.

Oleh Yuschuk got it exactly right with the OllyDbg. There were few releases – but they were properly tested and rock solid. From what I can see, x64dbg is going the other way:
compiles_ship_it

Frequent commits like “Fixed search for constant references”, “Fixed intermodular calls in module”, “Fixed FS/GS memory branch destinations” is not something you want to see in any software, let alone a debugger.

Well, it wouldn’t matter much, if there was some known-stable version I could put in my tool collection and use it anytime anywhere. But no, Mr. eXoDia thinks that “No more excuses to not update every day!” is a way to go. Instead of using tried-and-tested version, I should use a probably buggy and unstable one? Dafuq?

Conclusion

So, those are my 3 biggest complaints about x64dbg. I’d love to love x64dbg. I’d love to use x64dbg for everything. But right now I just can’t.

How about you?

01 Aug

Breaking B0rken ElGamal KeygenMe, part 2

In part 1 of the tutorial, I explained how badly initialized PRNG causes a serious problems and allows us to find the private key. In this part of tutorial, I’ll show how to use another weakness in crackme to find private key without using bruteforce.

Looking deeper in the keygenme

If you analyze serial check in more details, you’ll notice part of code that processes blacklisted names and keys:

It looks like we have 2 blacklisted names & corresponding serials. We can quickly patch crackme and verify that these names and serials really work.

But did you also notice that those serials look a lot alike? That’s weird.. Let’s take a closer look at them:

First part of the serial is exactly the same! Let’s go back to ElGamal basics and think a bit.

Revisiting ElGamal signing algorithm

I’m sure you already remember the algorithm from my previous blog post. But here it is again.. smile

To sign a message M, one would:

  • Make a hash of message, H(M). In this crackme, it’s SHA1 of the username
  • Generate a random number K where K<P-1
  • Calculate R=G^K (mod P)
  • Calculate S=(H(M)-RX)*K^(-1) (mod P-1)
  • The signature will be the pair C(R,S).

So, in our case, R is always the same. G and P are hardcoded in the crackme. That means.. K is always the same. And that doesn’t sound right! smile

Quick look in Wikipedia confirms that it’s a really bad idea:

The signer must be careful to choose a different k uniformly at random for each signature and to be certain that k, or even partial information about k, is not leaked. Otherwise, an attacker may be able to deduce the secret key x with reduced difficulty, perhaps enough to allow a practical attack. In particular, if two messages are sent using the same value of k and the same key, then an attacker can compute x directly.[1]

The problem of reusing k and the attack itself is explained in Stinson’s “Cryptography Theory And Practice”, pages 290-291. Unfortunately, normal person has no chance to understand that “explanation”.

Few Google searches later I found 2 writeups from Boston Key Party CTF 2015 which were slightly better:

So, let’s try to get something useful out of them.

Reused K and recovering private key

As I said earlier, cryptography is a dark magic – if you don’t spend years studying it, you can’t understand it. So, I just took the code from those CTF writeups and added few more comments to it. And no, I still don’t know why it works. smile

Conclusion

Wise man once said:

Knowledge is of two kinds. We know a subject ourselves, or we know where we can find information upon it.

You don’t need to be a crypto wizard to solve crackmes – you just need to know where to find the necessary information. But if you’re implementing cryptography in your software, better ask someone who understands those things.

Have fun!

22 Jul

Breaking B0rken ElGamal KeygenMe by SmilingWolf

Some weeks ago I found a nice keygenme on URET forum. The description looked interesting enough:

Yet another company is making wild claims! Your mission: prove that people shouldn’t trust companies promoting “revolutionary” crypto algos. Keygen this son of a crypto nightmare and write a DETAILED tutorial!

Rules:
1) The only acceptable solution is a keygen
2) No patching of course

It was not solved for few weeks, so I decided to take a look at it. smile

Crash-course in ElGamal signature scheme

I hate cryptography. It’s complex, it’s confusing and unless you’re prepared to study this field for years, you can’t really understand why stuff works this or that way.

So, here’s a short version, just enough to solve this keygenme. It’s based on the explanation in InfoSec Institute’s tutorial.

Key generation

  • Generate a random prime number P with chosen length.
  • Generate two random numbers, G and X, with G<P and X<P.
  • Calculate Y=G^X mod P.
  • Public key consists of 3 numbers: (P, G and Y), Private key is X.

All 3 numbers of public key are hardcoded in keygenme and are 256 bits long. We can find them in disassembly:

Private key X is.. well, private. smile Without it we can’t generate correct keys for a name of our choice. So, the challenge would be to recover the private key somehow.

Signing

To sign a message M, one would:

  • Make a hash of message, H(M). In this crackme, it’s SHA1 of the username
  • Generate a random number K where K<P-1
  • Calculate R=G^K (mod P)
  • Calculate S=(H(M)-RX)*K^(-1) (mod P-1)
  • The signature will be the pair C(R,S).

Verification

To verify a given pair C(R,S), one would:

  • Compute V1=G^M (mod P)
  • Compute V2=Y^R * R^S (mod P)
  • If V1==V2, the signature is valid

Breaking ElGamal

Strength of ElGamal algorithm lies in the Discrete Logarithm Problem (DLP). What it means is that easy to calculate Y=G^X mod P, but it’s bloody hard to find out X, if you know P, G and Y.

In the InfoSec Institute’s tutorial the author is using figugegl’s DLPTool to solve the problem for 128-bit integers. However, it’s not even possible to enter 256-bit integers in that tool. And the rest of the suggested tools can’t handle such large integers either.

So, we must find another way to break the keygenme. After all, it’s called “B0rken ElGamal”, so there must be a weakness somewhere!

Inside the keygenme

The keygenme itself is an application that generates ElGamal keys. It would be logical to assume that SmilingWolf used this same application to generate keys for the keygenme. So, let’s examine the application and see how ElGamal is implemented in it.

I’ll skip the boring “unpack modified UPX part”, as it’s been explained dozens of times already.

So, here’s the relevant code for key generation:

Looks legit, right? smile Well, not so fast.

Random numbers don’t just magically appear out of thin air. They are generated using random number generators. If the generators are flawed, the numbers are not really random. You can read a lot about such attacks on Wikipedia.

So, let’s examine random number generator used in this keygenme.

If you search for those constants, you’ll see that it’s a very standard PRNG which is not great but supposed to be reliable enough. You could bruteforce all 2^32 possibilities but it will take a long time. Another dead end?

Well, not really. Any PRNG must be initialized somehow, see random_seed variable above. So, let’s see how SmilingWolf is initializing his PRNG..

Hmm, that’s weird. Normally PRNG is initialized using rdtsc instruction or something even less predictable than that. And what exactly is arg_0? It’s a handle of the DialogBox – not random at all!

Finally, we’ve found a reason why this ElGamal implementation is broken! smile

Bruteforcing the private key

Now that we know the weakness, we can write a bruteforcer that will go through all possibilities of random seeds and generate all possible ElGamal keys. Once we generate a key that has the same P, G and Y as in the keygenme, we will also know the correct private key X. But generating all these numbers is a slow process!

Let’s look back at the code and see what we can optimize.

1) We don’t need to generate all the numbers. It’s enough if we find correct P – the rest of numbers will match automatically. So, let’s remove the rest of the code.

2) First 32 bits of 256 bit integer will almost certainly not change when searching for nearest prime. Therefore, we could get rid of bigint_find_nearest_prime call which is the slowest piece of code. This means we won’t be looking for number FE6D5B4400B30374A403F88CFBA3642435FB269AEC2BE5C8C2F331545EF37AB3, but any number starting with FE6D5B44…

3) Multiplication by 2. Another unnecessary step. Let’s just divide P by 2 and look for that number. Instead of looking for FE6D5B44…, we’ll look for 7F36ADA2…

4) Copying bigints. Unnecessary.

There’s not much left, I’m satisfied.

Now, how to bruteforce random seed? It’s a window handle. Window handles are always even. Also, they consist of 2 parts, high word and low word. Low word is the actual handle. It is almost never higher than 0x2000. High word is the “uniquifier” – just a counter. It’s usually quite small – on my PC it’s never larger than 0x800.

Taking all these assumptions into account, my pseudo-code for bruteforcer would look like this:

I implemented the bruteforcer in MASM32 with 95% of code ripped from keygenme and let it run. In less than 2 hours I had one (of the several possible) seed:

Once we know X, implementing keygen is a child’s play. Again, lots of ripped code and small UI around it. Problem solved!

Keygen for download: https://mega.nz/#!JtgVUDwb!xs3SgCDK3t5h3MabOfoNsDFYCj1mpHcI7GbQT-K1_RE

Further reading

InfoSec Institute’s tutorial about ElGamal: http://resources.infosecinstitute.com/breaking-software-protection-elgamal-signature-scheme/
Interactive webpage showing how ElGamal works with small numbers: https://asecuritysite.com/encryption/elgamal

Final thoughts

As SmilingWolf told me after solving his keygenme – there is another way how to break this keygenme. There’s one more thing broken in this implementation that makes generating keys really easy. smile So, if you’re interested, try to figure out what is it and how to abuse it.

08 Jun

What’s wrong with this file – ASLR is tricky!

I love magic tricks. My absolute favorites are “there’s nothing up my sleeve” kind of tricks. You can look at the equipment, you can examine magicians outfit, everything seems fine – yet the rabbit magically appears and disappears.

Here’s a similar reversing challenge for you: https://www.mediafire.com/?38evlc6gmyieskn

This EXE file contains relocations. It has all the necessary necessary flags in PE header. And it gets ASLR support in Windows 10, as you can see in picture:
win10_has_aslr
But on Windows 7/8.1 this poor executable will be always loaded at it’s preferred imagebase 0x400000, and doesn’t get ASLR support:
win7_no_aslr
Can you figure out what’s so special about it? smile

I will provide the correct answer in one week. Or you can provide your opinions in comments. Extra respect awarded for detailed answers and explaining how you figured that out. Extra extra respect if you knew the answer even before looking at the executable. smile

20 May

Beautiful code

After making quite a few unpackers and other RE-related tools, publishing sources for them and having to maintain and bugfix them, all I can say is: “Read this. Remember this. Worship this.

All code is born ugly.

It starts disorganized and inconsistent, with overlaps and redundancies and gaps.

We begin working it into an imperfect solution for an often poorly defined problem.

As we start building up like clay, a solution starts taking form. The feedback guides us in moving, removing and adding material. It allows us to add and remove details. We learn from our mistakes.

Thank you, Dennis, you made my day so much better.

11 Mar

About .NET, googling and lazy programmers.

Delphi fail. .NET win.

Recently, several people sent me bug reports where my EnigmaVB unpacker failed to extract files. In all cases, the problem was caused by really large files, like 3.5GB in size. So, what’s causing the problem?

EnigmaVB unpacker is a 32bit application written in Delphi. And Delphi streams are retarded. They look like they’ve been written in 1990s and were never updated. TMemoryStream uses a continuous memory range, so it can never support anything larger than 2GB. TFileStream internally uses longint, so it doesn’t properly support files larger than 2GB. WTF?

So, I have two choices. I can either make a custom stream class in Delphi, or I can pick another framework and rewrite my tool to use that.

I’m not a programmer, I’m a reverser. I don’t want to spend my time developing custom stream classes. I’d very much rather use this time breaking some code. So, say hello to .NET framework – my next version of EnigmaVB unpacker will be made in C#.. smile

Am I a programmer or a googler?

While researching all the Delphi limitations and possible workarounds, I ran into this great article by Scott Hanselman. Reading both the post and the comments made me think a lot.

Does using Google to solve your programming tasks makes you less of a programmer? I don’t think so.

In fact, I’m just lazy. Most people are. Why should I spend 30 minutes remembering basic algorithms for converting string to hex, if Google query can solve it in 10 seconds? Why reinvent the wheel and write CRC calculation from scratch? I’ll just open StackOverflow and have a solution that’s already tried and tested. It doesn’t mean I can’t do those boring tasks – I just don’t want to.

How about you? Would you be able to write some tools without using Google and StackOverflow?

16 Feb

Updated Meltdown and EnigmaVB Unpacker

About Error Messages

Users can’t read.

Or maybe they don’t want to read. I don’t know.

But one thing I know for sure – you must make your tools foolproof. If your tool is showing an error message, make sure even your grandma could understand it. Otherwise you’ll be getting lots and lots of invalid bug reports.

For example, this is the error message my EnigmaVB unpacker used to show (as reported by ho3ein at Tuts4You):
enigmavb_error_message

It seemed to be very clear to me. First, tool tells user all the versions of Enigma Virtual Box it supports. Then tool explains that it expects to see a PE section with a name “.enigma2” but it found section with a name “.rsrc” instead. To me it’s absolutely clear what happened: this file is not protected with Enigma Virtual Box (or it’s hacked).

But you won’t believe how many times this gets reported as a bug.

There was a similar problem with Meltdown. It clearly stated which versions of DeepFreeze it supports. Then it printed the detected DeepFreeze version. However, the error message didn’t explicitly say “This version of DeepFreeze is not supported”, it said “DeviceIoControl failed.” It makes perfect sense from developer’s point of view, but apparently is very confusing for users.

So, here are improved versions of my tools, fixing the error messages and some other stuff..

Improved EnigmaVB Unpacker

First of all, I fixed the error message. I also added detection and tested compatibility with the latest EnigmaVB v7.40. Hopefully, this will make users happier and less confused. smile
EnigmaVBUnpacker_v034

Improved Meltdown

Meltdown 1.7 fixes confusing error message with DeepFreeze Standard v8.x. Thanks to Alexander for reporting it.

I also took a closer look at DeepFreeze Enterprise versions and found a way to make Meltdown more user friendly. If DeepFreeze Enterprise v7.20+ is detected, Meltdown will get OTP Token automatically and immediately generate correct password.

meltdown17

Download links

Enigma Virtual Box Unpacker: Please get latest version from this post
Meltdown v1.7: https://www.mediafire.com/?b0bamd3t2d6bbkq

05 Nov

Keygen templates in Visual Studio

I’m lazy and I hate doing the same tasks over and over again. Making UI for my crackme solutions is one of such tasks. It always goes like this: open Visual Studio, create new Windows Forms project in C#, drop 2 labels, 2 edit boxes and one button on the form. Set label texts to “Name” and “Serial”, set button title to “Generate..”, set the project icon, etc., etc..

There must be a better way!

..and it’s certainly not the way Blue Indian did his keygen template:

To build this template on your own, open the solution in Visual studio, comment out the calls for uFMOD and implement your own logic, after successful build of keygen, close the Visual studio, open the Form_Main.cs file in any text editor and uncomment those two calls to uFMod, save it. Now double click on the build.bat file to built it finally.

-To change the ICON and XM tune, edit the mini.res (resource file) with any resource editor like Restorator or any of your choice.

Open this, delete those, compile that, and what? I’m already confused, sorry.

Introducing Visual Studio project templates

I’m sure you know that when you click “New project” in Visual Studio, you’re presented with number of choices, like “Windows Forms Application”, “Console Application”, “Class Library” and so on. All these are project templates that are installed by default.

They provide all the files that are required for a particular project type, include standard assembly references, and set default project properties and compiler options. Hmm, that’s exactly what I needed! smile

This article at MSDN nicely explains that project template is simply a ZIP file that contains all the necessary files and a special .vstemplate file. This .vstemplate file is an XML file containing metadata Visual Studio needs to display the template in the “New Project” dialog.

Let’s try to put it all together.

Making simple keygen template

Making a new template is actually very easy. You take an existing Visual Studio project, replace project-specific strings with template parameters and press File->Export Template.

Here is my keygen for Mr. eXoDia’s simple crackme:
keygen_template1
Obviously, template should not contain code for specific crackme. Let’s change that to something trivial and mark as FIXME:
keygen_template2
Now I need to remove all references to crackme name. I will replace them with template parameter $safeprojectname$ in all files. After this change, project won’t compile anymore, so you need to be extra careful when changing stuff!
keygen_template3
Hardcoding year in the (c) string is not a good idea because I want to use this template in year 2016 as well:
keygen_template4
Now I just need to update AssemblyInfo.cs to make sure each project has correct name, (c) and GUIDs:
keygen_template5
Did it work? Let’s see… File->Export Template, follow the wizard and…

It works. Kinda. The created template still has quite a few references to Mr eXodia’s crackme, I’ll need to modify project and solution files manually. Unzip the template, fix the files in text editor and ZIP them back. And now it works!

Few more cosmetic fixes (like using $projectname$ where possible), using $if$ and $targetframeworkversion$ to target all .NET framework versions, better namespace names and we have a template that’s actually useful.

Download here: https://www.mediafire.com/?sx1i5ba1uijjkii

It’s not particularly pretty but that’s pretty much what I’ve been using for 2+ years now – and hopefully it can inspire you to do something similar with your own code. wink

Further reading

Reason→Code→Example : Creating Visual Studio project templates
Rebuilding template cache
How to: Manually Create Project Templates
How to: Create Multi-Project Templates

29 Jun

Linking OMF files with Delphi

Continuing the discussion about Delphi compiler and the object files.

Here is the OMF file template I made for 010 Editor: https://www.mediafire.com/?bkpbkjvgen7ubz1
omf_parser

Please note, it is not a full-featured implementation of OMF specification. I only implemented all OMF file records that are processed by Delphi 2007 compiler. So, next time you have a cryptic compiler error while trying to link OMF file in Delphi, you can take a look into your OBJ file and make an educated guess what’s causing the problem.

TL;DR version

In 95+% of cases you will encounter OBJ file that has unsupported segment name in SEGDEF record. And it’s a simple fix, too – you just need to use objconv.exe by Agner Fog and use -nr option to rename the offending segment. Something like this:

Next possible issue is exceeding the number of EXTDEF or LNAMES records – this can happen if you’re trying to convert a really large DLL file into OBJ file.

Finally, your OBJ file might contain some record type which is not supported by Delphi compiler at all. I’m not aware of a simple way to fix it, I would try using 010Editor and OMF template to remove the entire record.

If your problem is not caused from any of the above issues, please feel free to drop me a note – I’ll be happy to look into it.

Known limitations of Delphi compiler

This is a list of limitations I was able to compile and/or confirm. Some of them come from Embarcadero official notes and the rest I obtained by analyzing dcc32.exe.

SEGDEF (98H, 99H)

  • Not more than 10 segments – if number of segments exceeds 10, buffer overrun will probably happen.
  • Segments must be 32bits. Will cause “E2215 16-Bit segment encountered in object file ‘%s'”
  • Segment name must be one of (case insensitive):
    • Code segments: “CODE”, “CSEG”, “_TEXT”
    • Constant data segments: “CONST”, “_DATA”
    • Read-write data segments: “DATA”, “DSEG”, “_BSS”

    Segment with any other name will be ignored.

LNAMES (96H)

Not more than 50 local names in LNAMES records – will cause “E2045 Bad object file format: ‘%s'” error.

EXTDEF (8CH)

Not more than 255 external symbols – will cause “E2045 Bad object file format: ‘%s'”
Certain EXTDEF records can also cause “E2068 Illegal reference to symbol ‘%s’ in object file ‘%s'” and “E2045 Bad object file format: ‘%s'”

PUBDEF (90H, 91H)

Can cause “E2045 Bad object file format: ‘%s'” and “F2084 Internal Error: %s%d”

LEDATA (A0H, A1H)

Embarcadero says that “LEDATA and LIDATA records must be in offset order” – I am not really sure what that means. Can cause “E2045 Bad object file format: ‘%s'”

LIDATA (A2H, A3H)

Embarcadero says that “LEDATA and LIDATA records must be in offset order” – I am not really sure what that means. Can cause “E2045 Bad object file format: ‘%s'”

FIXUPP (9CH)

This type of record is unsupported, will cause immediate error “E2103 16-Bit fixup encountered in object file ‘%s'”

FIXUPP (9DH)

Embarcadero documentation says:

  • No THREAD subrecords are supported in FIXU32 records
  • Only segment and self relative fixups
  • Target of a fixup must be a segment, a group or an EXTDEF

Again I’m not sure what they mean. But there are lots of checks that can cause “E2045 Bad object file format: ‘%s'”

THEADR (80H)

Accepted by compiler, but no real checks are performed.

LINNUM (94H, 95H)

Accepted by compiler, but no real checks are performed.

MODEND (8AH, 8BH)

Accepted by compiler, but no real checks are performed.

COMMENT (88H) and GRPDEF (9AH)

Ignored by compiler.

That’s the end of the list. Any other entry type will cause immediate error “E2045 Bad object file format: ‘%s'” smile

Useful links

My OMF file template for 010Editor: https://www.mediafire.com/?bkpbkjvgen7ubz1
OMF file format specification.
The Borland Developer’s Technical Guide
Objconv.exe by Agner Fog
Manual for objconv.exe