AnalyzePDF - Bringing the Dirt Up to the Surface

This has been ported over to my GitHub site and is not longer being maintained here. For any issues, comments or updates head here. 

What is that thing they call a PDF?

The Portable Document Format (PDF) is an old format ... it was created by Adobe back in 1993 as an open standard but wasn't officially released as an open standard (SIO 32000-1) until 2008 - right @nullandnull ?  I can't take credit for the nickname that I call it today, Payload Delivery Format, but I think it's clever and applicable enough to mention.  I did a lot of painful reading through the PDF specifications in the past and if you happen to do the same I'm sure you'll also have a lot of "hm, that's interesting" thoughts as well as many "wtf, why?" thoughts.  I truly encourage you to go out and do the same... it's a great way to learn about the internals of something, what to expect and what would be abnormal.  The PDF has become a defacto for transferring files, presentations, whitepapers etc.

<rant> How about we stop releasing research/whitepapers about PDF 0-days/exploits via a PDF file... seems a bit backwards</rant>

We've all had those instances where you wonder if that file is malicious or benign ... do you trust the sender or was it downloaded from the Internet?   Do you open it or not?  We might be a bit more paranoid than most people when it comes to this type of thing and but since they're so common they're still a reliable means for a delivery method by malicious actors.  As the PDF contains many 'features', these features often turn into 'vulnerabilities' (Do we really need to embed an exe into our PDF? or play a SWF game?).  Good thing it doesn't contain any vulnerabilities, right? (to be fair, the sandboxed versions and other security controls these days have helped significantly)

http://www.cvedetails.com/product/497/Adobe-Acrobat-Reader.html?vendor_id=53

What does a PDF consist of?

In its most basic format, a PDF consists of four components: header, body, cross-reference table (Xref) and trailer:

(sick M$ Paint skillz, I know)

If we create a simple PDF (this example only contains a single word in it) we can see a better idea of the contents we'd expect to see:

 What else is out there?

Since PDF files are so common these days there's no shortage of tools to rip them apart and analyze them.  Some of the information contained in this post and within the code I'm releasing may be an overlap of others out there but that's mainly because the results of our research produced similar results or our minds think alike...I'm not going to touch on every tool out there but there are some that are worth mentioning as I either still use them in my analysis process or some of their functionality/lack of functionality is what sparked me to write AnalyzePDF.  By mentioning the tools below my intentions aren't to downplay them and/or their ability to analyze PDF's but rather helping to show reasons I ended up doing what I did.

pdfid/pdf-parser

Didier Stevens created some of the first analysis tools in this space, which I'm sure you're already aware of.  Since they're bundled into distros like BackTrack/REMnux already they seem like good candidates to leverage for this task.  Why recreate something if it's already out there?  Like some of the other tools, it parses the file structure and presents the data to you... but it's up to you to be able to interpret that data.  Because these tools are commonly available on distros and get the job done I decided they were the best to wrap around.

Did you know that pdfid has a lot more capability/features that most aren't aware of?  If you run it with the (-h) switch you'll see some other useful options such as the (-e) which display extra information. Of particular note here is the mention of "%%EOF", "After last %%EOF", create/mod dates and the entropy calculations.  During my data gathering I encountered a few hiccups that I hadn't previously experienced.  This is expected as I was testing a large data set of who knows what kind of PDF's.  Again, I'm not noting these to put down anyone's tools but I feel it's important to be aware of what the capabilities and limitations of something are - and also in case anyone else runs into something similar so they have a reference.  Because of some of these, I am including a slightly modified version of pdfid as well.  I haven't tested if the newer version fixed anything so I'd rather give the files that I know work with it for everyone.

• I first experienced a similar error as mentioned here when using the (-e) option on a few files (e.g. - cbf76a32de0738fea7073b3d4b3f1d60).  It appears it doesn't count multiple '%%EOF's since if the '%%EOF' is the last thing in the file without a '/r' or '/n' behind it, it doesn't  seem to count it.
• I've had cases where the '/Pages' count was incorrect - there were (15) PDF's that showed '0' pages during my tests.  One way I tried to get around this was to use the (-a) option and test between the '/Page' and '/Pages/ values. (e.g. - ac0487e8eae9b2323d4304eaa4a2fdfce4c94131)
• There were times when the number of characters after the last '%%EOF' were incorrect
• Won't flag on JavaScript if it's written like "<script contentType="application/x-javascript">" (e.g - cbf76a32de0738fea7073b3d4b3f1d60) :

peepdf

Peepdf has gone through some great development over the course of me using it and definitely provides some great features to aid in your analysis process.  It has some intelligence built into it to flag on things and also allows one to decode things like JavaScript from the current shell.  Even though it has a batch/automated mode to it, it still feels like more of a tool that I want to use to analyze a single PDF at a time and dig deep into the files internals.

• Originally, this tool didn't look match keywords if they had spaces after them but it was a quick and easy fix... glad this testing could help improve another users work.

PDFStreamDumper

PDFStreamDumper is a great tool with many sweet features but it has its uses and limitations like all things.  It's a GUI and built for analysis on Windows systems which is fine but it's power comes from analyzing a single PDF at a time - and again, it's still mostly a manual process.

pdfxray/pdfxray_lite

Pdfxray was originally an online tool but Brandon created a lite version so it could be included in REMnux (used to be publicly accessible but at the time of writing this looks like that might have changed).  If you look back at some of Brandon's work historically he's also done a lot in this space as well and since I encountered some issues with other tools and noticed he did as well in the past I know he's definitely dug deep and used that knowledge for his tools.  Pdfxray_lite has the ability to query VirusTotal for the file's hash and produce a nice HTML report of the files structure - which is great if you want to include that into an overall report but again this requires the user to interpret the parsed data

pdfcop

Pdfcop is part of the Origami framework.  There're some really cool tools within this framework but I liked the idea of analyzing a PDF file and alerting on badness.  This particular tool in the framework has that ability, however, I noticed that if it flagged on one cause then it wouldn't continue analyzing the rest of the file for other things of interest (e.g. - I've had it close the file our right away if there was an invalid Xref without looking at anything else.  This is because PDF's are read from the bottom up meaning their Xref tables are first read in order to determine where to go next).  I can see the argument of saying why continue to analyze the file if it already was flagged bad but I feel like that's too much of tunnel vision for me.  I personally prefer to know more than less...especially if I want to do trending/stats/analytics.

So why create something new?

While there are a wealth of PDF analysis tools these days, there was a noticeable gap of tools that have some intelligence built into them in order to help automate certain checks or alert on badness.  In fairness, some (try to) detect exploits based on keywords or flag suspicious objects based on their contents/names but that's generally the extent of it.  I use a lot of those above mentioned tools when I'm in the situation where I'm handed a file and someone wants to know if it's malicious or not... but what about when I'm not around?  What if I'm focused/dedicated to something else at the moment?  What if there's wayyyy too many files for me to manually go through each one?  Those are the kinds of questions I had to address and as a result I felt I needed to create something new.  Not necessarily write something from scratch... I mean why waste that time if I can leverage other things out there and tweak them to fit my needs?  

Thought Process

What do people typically do when trying to determine if a PDF file is benign or malicious?  Maybe scan it with A/V and hope something triggers, run it through a sandbox and hope the right conditions are met to trigger or take them one at a time through one of the above mentioned tools?  They're all fine work flows but what if you discover something unique or come across it enough times to create a signature/rule out of so you can trigger on it in the future?  We tend to have a lot to remember so doing the analysis one offs may result in us forgetting something that we previously discovered.  Additionally, this doesn't scale too great in the sense that everyone on your team might not have the same knowledge that you do... so we need some consistency/intelligence built in to try and compensate for these things.<

 I felt it was better to use the characteristics of a malicious file (either known or observed from combinations of within malicious files) to eval what would indicate a malicious file.  Instead of just adding points for every questionable attribute observed. e.g. - instead of adding a point for being a one page PDF, make a condition to say if you see an invalid xref and a one page PDF then give it a score of X.  This makes the conditions more accurate in my eyes; since, for example:

1. A single paged PDF by itself isn't malicious but if it also contains other things of question then it should have a heavier weight of being malicious.  
2. Another example is JavaScript within a PDF.  While statistics show JavaScript within a PDF are a high indicator that it's malicious, there're still legitimate reasons for JavaScript to be within a PDF (e.g. - to calculate a purchase order form or verify that you correctly entered all the required information the PDF requires).

Gathering Stats

At the time I was performing my PDF research and determining how I wanted to tackle this task I wasn't really aware of machine learning.  I feel this would be a better path to take in the future but the way I gathered my stats/data was in a similar (less automated/cool AI) way.  There's no shortage of PDF's out there which is good for us as it can help us to determine what's normal, malicious, or questionable and leverage that intelligence within a tool.

If you need some PDF's to gather some stats on, contagio has a pretty big bundle to help get you started.  Another resource is Govdocs from Digital Corpora ... or a simple Google dork.

Note : Spidering/downloading these will give you files but they still need to be classified as good/bad for initial testing).  Be aware that you're going to come across files that someone may mark as good but it actually shows signs of badness... always interesting to detect these types of things during testing!

Stat Gathering Process

So now that I have a large set of files, what do I do now?  I can't just rely on their file extensions or someone else saying they're malicious or benign so how about something like this:

1. Verify it's a PDF file.  
• When reading through the PDF specs I noticed that the PDF header can be within the first 1024 bytes of the file as stated in ""3.4.1, 'File Header' of Appendix H - ' Acrobat viewers require only that the header appear somewhere within the first 1024 bytes of the file.'"... that's a long way down compared to the traditional header which is usually  right in the beginning of a file.  So what's that mean for us?  Well if we rely solely on something like file or TRiD they _might_ not properly identify/classify a PDF that has the header that far into the file as most only look within the first 8 bytes (unfair example is from corkami).  We can compensate for this within our code/create a YARA rule etc.... you don't believe me you say?  Fair enough, I don't believe things unless I try them myself either:

The file to the left is properly identified as a PDF file but when I created a copy of it and modified it so the header was a bit lower, the tools failed.  The PDF on the right is still in accordance with the PDF specs and PDF viewers will still open it (as shown)... so this needs to be taken into consideration.





Get rid of duplicates (based on SHA256 hash) for both files in the same category (clean vs. dirty) then again via the entire data set afterwards to make sure there're no duplicates between the clean and dirty sets.
Run pdfid & pdfinfo over the file to parse out their data.  


• These two are already included in REMnux so I leveraged them. You can modify them to other tools but this made it flexible for me and I knew the tool would work when run on this distro; pdfinfo parsed some of the data better during tests so getting the best of both of them seemed like the best approach.



Run scans for low hanging fruit/know badness with local A/V||YARA
Now that we have a more accurate data set classified:




Are all PDFs classified as benign really benign?
Are all PDFs classified as malicious really malicious? 

Stats

Files analyzed (no duplicates found between clean & dirty):

Class	Type	Count
Dirty	Pre-Dup	22,342
Dirty	Post-Dup	11,147
Clean	Pre-Dup	2,530
Dirty	Post-Dup	2,529
Total Files Analyzed:		13,676

I've collected more than enough data to put together a paper or presentation but I feel that's been played out already so if you want more than what's outlined here just ping me.  Instead of dragging this post on for a while showing each and every stat that was pulled I feel it might be more useful to show a high level comparison of what was detected the most in each set and some anomalies.

Ah-Ha's

• None of the clean files had incorrect file headers/versions
• There wasn't a single keyword/attribute parsed from the clean files that covered more than 4.55% of it's entire data set class.  This helps show the uniqueness of these files vs. malicious actors reusing things.
• The dates within the clean files were generally unique while the date fields on the dirty files were more clustered together - again, reuse?
• None of the values for the keywords/attributes of the clean files were flagged as trying to be obfuscated by pdfid
• Clean files never had '/Colors > 2^24' above 0 while some dirty files did 
• Rarely did a clean file have a high count of JavaScript in it while dirty files ranged from 5-149 occurrences per file
• '/JBIG2Decode' was never above '0' in any clean file
• '/Launch' wasn't used much in either of the data sets but still more common in the dirty ones
• Dirty files have far more characters after the last %%EOF (starting from 300+ characters is a good check)
• Single page PDF's have a higher likelihood of being malicious - no duh
• '/OpenAction' is far more common in malicious files

YARA signatures

I've also included some PDF YARA rules that I've created as a separate file so you can use those to get started.  YARA isn't really required but I'm making it that way for the time being because it's helpful... so I have the default rules location pointing to REMnux's copy of MACB's rules unless otherwise specified.

Clean data set:

Dirty data set:

Signatures that triggered across both data sets:

Cool... so we know we have some rules that work well and others that might need adjusting, but they still help!

What to look for

So we have some data to go off of... what are some additional things we can take away from all of this and incorporate into our analysis tool so we don't forget about them and/or stop repetitive steps?

1. Header
• In addition to being after the first 8 bytes I found it useful to look at the specific version within the header.  This should normally look like "%PDF-M.N." where M.N is the Major/Minor version .. however, the above mentioned 'low header' needs to be looked for as well.
  
  Knowing this we can look for invalid PDF version numbers or digging deeper we can correlate the PDF's features/elements to the version number and flag on mismatches. Here're some examples of what I mean, and more reasons why reading those dry specs are useful:
• If FlateDecode was introduced in v1.2 then it shouldn't be in any version below
• If JavaScript and EmbeddedFiles were introduced in v1.3 then they shouldn't be in any version below
• If JBIG2 was introduced in v1.4 then it shouldn't be in any version below
2. Body
• This is where all of the data is (supposed to be) stored; objects (strings, names, streams, images etc.).  So what kinds of semi-intelligent things can we do here?
• Look for object/stream mismatches.  e.g - Indirect Objects must be represented by 'obj' and 'endobj' so if the number of 'obj' is different than the number of  'endobj' mentions then it might be something of interest
• Are there any questionable features/elements within the PDF? 
• JavaScript doesn't immediately make the file malicious as mentioned earlier, however, it's found in ~90% of malicious PDF's based on others and my own research.
• '/RichMedia'  - indicates the use of Flash (could be leveraged for heap sprays)
• '/AA', '/OpenAction', '/AcroForm' - indicate that an automatic action is to be performed (often used to execute JavaScript)
• '/JBIG2Decode', '/Colors' - could indicate the use of vulnerable filters; Based on the data above maybe we should look for colors with a value greater than 2^24
• '/Launch', '/URL', '/Action', '/F', '/GoToE', '/GoToR' - opening external programs, places to visit and redirection games
• Obfuscation
• Multiple filters ('/FlateDecode', '/ASCIIHexDecode', '/ASCII85Decode', '/LZWDecode', '/RunLengthDecode')
•  The streams within a PDF file may have filters applied to them (usually for compressing/encoding the data).  While this is common, it's not common within benign PDF files to have multiple filters applied.  This behavior is commonly associated with malicious files to try and thwart A/V detection by making them work harder.
• Separating code over multiple objects
• Placing code in places it shouldn't be (e.g. - Author, Keywords etc.)
• White space randomization
• Comment randomization
• Variable name randomization
• String randomization
• Function name randomization
• Integer obfuscation
• Block randomization
• Any suspicious keywords that could mean something malicious when seen with others?
•  eval, array, String.fromCharCode, getAnnots, getPageNumWords, getPageNthWords, this.info, unescape, %u9090
3. Xref
The first object has an ID 0 and always contains one entry with generation number 65535. This is at the head of the list of free objects (note the letter ‘f’ that means free). The last object in the cross reference table uses the generation number 0.

Translation please?  Take a look a the following Xref:


Knowing how it's supposed to look we can search for Xrefs that don't adhere to this structure.

4. Trailer
• Provides the offset of the Xref (startxref)
• Contains the EOF, which is supposed to be a single line with "%%EOF" to mark the end of the trailer/document.  Each trailer will be terminated by these characters and should also contain the '/Prev' entry which will point to the previous Xref.
• Any updates to the PDF usually result in appending additional elements to the end of the file
  
  This makes it pretty easy to determine PDF's with multiple updates or additional characters after what's supposed to be the EOF
5. Misc.
• Creation dates (both format and if a particular one is known to be used)
• Title
• Author
• Producer
• Creator
• Page count

The Code

So what now?  We have plenty of data to go on - some previously known, but some extremely new and helpful.  It's one thing to know that most files with JavaScript or that are (1) page have a higher tendency of being malicious... but what about some of the other characteristics of these files?  By themselves, a single keyword/attribute might not stick out that much but what happens when you start to combine them together?  Welp, hang on because we're going to put this all together.

File Identification

In order to account for the header issue, I decided the tool itself would look within the first 1024 bytes instead of relying on other file identification tools:

Another way, so this could be detected whether this tool was used or not, was to create a YARA rule such as:

Wrap pdfinfo

Through my testing I found this tool to be more reliable in some areas as opposed to pdfid such as:

• Determining if there're any Xref errors produced when trying to read the PDF
• Look for any unterminated hex strings etc.
• Detecting EOF errors

Wrap pdfid

• Read the header.  *pdfid will show exactly what's there and not try to convert it*
• _attempt_ to determine the number of pages
• Look for object/stream mismatches
• Not only look for JavaScript but also determine if there's an abnormally high amount
• Look for other suspicious/commonly used elements for malicious purposes (AcroForm, OpenAction, AdditionalAction, Launch, Embedded files etc.)
• Look for data after EOF
• Calculate a few different entropy scores

Next, perform some automagical checks and hold on to the results for later calculations.

Scan with YARA

While there are some pre-populated conditions that score a ranking built into the tool already, the ability to add/modify your own is extremely easy.  Additionally, since I'm a big fan of YARA I incorporated it into this as well.  There're many benefits of this such as being able to write a rule for header evasion, version number mismatching to elements or even flagging on known malicious authors or producers.  The biggest strength, however, is the ability to add a 'weight' field in the meta section of the YARA rules.  What this does is allow the user to determine how good of a rule it is and if the rule triggers on the PDF, then hold on to its weighted value and incorporate it later in the overall calculation process which might increase it's maliciousness score.  Here's what the YARA parsing looks like when checking the meta field:

And here's another YARA rule with that section highlighted for those who aren't sure what I'm talking about:

If the (-m) option is supplied then if _any_ YARA rule triggers on the PDF file it will be moved to another directory of your choosing.  This is important to note because one of your rules may hit on the file but it may not be displayed in the output, especially if it doesn't have a weight field.

Once the analysis has completed the calculation process starts.  This is two phase -

1. Anything noted from pdfino and pdfid are evaluated against some pre-determined combinations I configured.  These are easy enough to modify as needed but they've been very reliable in my testing...but hey, things change!  Instead of moving on once one of the combination sets is met I allow the scoring to go through each one and add the additional points to the overall score, if warranted.  This allows several 'smaller' things to bundle up into something of interest rather than passing them up individually.
2. Any YARA rule that triggered on the PDF file has it's weighted value parsed from the rule and added to the overall score.  This helps bump up a files score or immediately flag it as suspicious if you have a rule you really want to alert on.

So what's it look like in action?  Here's a picture I tweeted a little while back of it analyzing a PDF exploiting CVE-2013-0640 :

Download

I've had this code for quite a while and haven't gotten around to writing up a post to release it with but after reading a former coworkers blog post last night I realized it was time to just write something up and get this out there as there are still people asking for something that employs some of the capabilities (e.g. - weight ranking).  Is this 100% right all the time? No... let's be real.  I've come across situations where a file that was benign was flagged as malicious based on its characteristics and that's going to happen from time to time.  Not all PDF creators adhere to the required specifications and some users think it's fun to embed or add things to PDF's when it's not necessary.  What this helps to do is give a higher ranking to files that require closer attention or help someone determine if they should open a file right away vs. send it to someone else for analysis (e.g. - deploy something like this on a web server somewhere and let the user upload their questionable file to is and get back a "yes it's ok -or- no, sending it for analysis".

AnalyzePDF can be downloaded on my github

Further Reading

• Research papers (one | two | three) [PDF]
• PDFTricks
• PDF Overview

NoMoreXOR

This has been ported over to my GitHub site and is not longer being maintained here. For any issues, comments or updates head here.

Update 04/09/2013 - NoMoreXOR is now included in REMnux as of version 4.

Have you ever been faced with a file that was XOR'ed with a 256 byte key? While it may not be the most common length for an XOR key, it's still something that has popped up enough over the last few months to make it on my to-do list.  If you take a look at first the two links mentioned above you'll see they both include some in-house tool(s) which do some magic and provide you with the XOR key.  Even though they both state that at some point their tools will be released, that doesn't help me now.

Most of the tools I came across can handle single byte - four byte XOR keys no problem (xortool, xortools, XORBruteForcer, xorsearch etc.) but other than that I didn't notice any that would handle (or actually work) with a large XOR key besides for (okteta, converter and cryptam_unxor).

I noticed Cryptam's online document analysis tool had the ability to do this as well so I sent them a few questions on their process and received a quick, informative response which pointed me to a post on their site.  Within the post/email they said that they don't perform any bruteforcing on the XOR key but rather perform cryptanalysis and then brute force the ROL1-7 (if present).  As shown in the dispersion graphs they provide, they appear to essentially be looking for high frequencies of repetitive data then using whatever appears the most to test as the key(s).

So how do you know if the file is XOR'ed with a 256 byte key in the first place?  Well... you could always try to reverse it but you may also be lucky enough to have some YARA rules which have some pre-calculated rules to help aid in this situation.  A good start would be to look at MACB's xotrools (previously linked) and also consider what it is you might want to look for (i.e. - "This program cannot be run") and XOR it with some permutations.

Manual process

If we open that file within a hex editor and go to the offset flagged (0x25C8) we'll see what is supposedly "This program cannot be run" = 26 bytes :

If we take that original file and covert it to hex we'll essentially just get a big hex blob:

...but that hex blob helps to try and guess the XOR key:

From my initial tests, the XOR key has always been in the top returned results, but even if you're having some difficulties for whatever reason you can always modify the code to fit your needs - gotta love that.

So if we now try to unxor the original file with the first guessed XOR key (remember XOR is symmetric) hopefully we'll get the original content that was XOR'ed:

After the original file was unxored and scanned with YARA we see that it was flagged for having an embedded EXE within it (this rule can be found within MACB's capabilities.yara file) so it looks like it worked.

Now while all this hex may look like a bunch of garbage at times, the human eye is very good at recognizing patterns - and when you look more and more at things like this you'll start to recognize them.  Do you recall the YARA hit that triggered? It stated that the XOR key was incremented.  What this means is that each byte is being XOR'ed with the next byte in an incremental fashion until it wraps back around to the beginning.  That may be confusing the grasp at first so lets visualize it by breaking down the previously found 256 byte XOR key in its' respective order:

868788898a8b8c8d8e8f
909192939495969798999
a9b9c9d9e9f
a0a1a2a3a4a5a6a7a8a9
aaabacadaeaf
b0b1b2b3b4b5b6b7b8b9
babbbcbdbebf
c0c1c2c3c4c5c6c7c8c9
cacbcccdcecf
d0d1d2d3d4d5d6d7d8d9
dadbdcdddedfe
0e1e2e3e4e5e6e7e8e9
eaebecedeeef
f0f1f2f3f4f5f6f7f8f9
fafbfcfdfeff
000102030405060708090
a0b0c0d0e0f
10111213141516171819
1a1b1c1d1e1f
20212223242526272829
2a2b2c2d2e2f
30313233343536373839
3a3b3c3d3e3f
40414243444546474849
4a4b4c4d4e4f
50515253545556575859
5a5b5c5d5e5f
60616263646566676869
6a6b6c6d6e6f
70717273747576777879
7a7b7c7d7e7f
808182838485

As you see, it started with 86 and looped all the way around till it reached 85 - you should also notice the patterns on each line.  This is just an example of incremental/decremental XOR (not as commonly observed in my testing but useful to be aware of) but it's useful to know because it's quite easy to spot if you look at the original file in a hex editor again:

... and that's a pattern that was observed repeating ~56 times.

Automated process

So now we can kind of put together a process flow of what we want to do:

1. Convert the original, XOR'ed file to hex
2. Conduct some slight frequency analysis of the newly created hex file and look for the most common characters as well as the most commonly observed hex chunks.  
1. The first part may help in determining if there's an embedded PE file (usually a lot of \x00's) or possibly help deduce if certain bytes should be skipped.  
2. The latter essentially reads 512 bytes at a time, stores it and continues till the end of the file.  Once complete it does some simple checking to try and weave out meaningless possible keys then presents the top five most observed 512 bytes or characters in this sense  (i.e. - 512 characters = 1 possible 256 byte key(s))
3. For each possible XOR key guessed from the previous step, XOR (the entire file for right now) the original file, save it to a new file and scan it with YARA.  
1. I chose to perform YARA scans here to help determine the likelihood that the key used was correct - you may choose to implement something else such as just a check for an embedded PE file etc.  If there are YARA hits then I stop attempting the other possible XOR keys (if any other were still to be processed) and assume the previous XOR key was the correct one.

* If you stick with the YARA scanning, it will continue to process all of the possible key(s) it outlined as the top, in terms of frequency, so your YARA rules should include something that might be present in the original XOR'ed file.  If not, you might already have the correct XOR key but aren't aware.  Embedded exe's are a good start to look for since they're common - but remember if we XOR the entire file at once instead of a specific section that you might find the embedded content but that doesn't mean the original file will be readable afterwards (i.e - won't be a Word document anymore since it was XOR'ed) 

Let's try out that process flow in a more automated way (on a new file):

As you can see, it worked like a charm :)

As always, I'm sure there's a better way to code some of the stuff I did but hey, it works for me at the moment.  There's a to-do list of things that I want to further implement into this tool, some of which is already included in other tools.  I've been asked before how this tools will work with smaller XOR keys and that's up to you to test and tell me - I created this in order to tackle the problem solely of the 256 byte key files I was observing so I'd recommend using one of the earlier mentioned tools for that situation, at least for the time being.

Example To-do's:

• ROL(1-7)/ROT(1-25) - either brute forcing or via YARA scans
• Add ability to skip \x00 & other chosen bytes (ref)
• more is outlined within the file....

Download

NoMoreXOR can be found on my github

Customizing cuckoo to fit your needs

This has been ported over to my GitHub site and is not longer being maintained here. For any issues, comments or updates head here.

With the talk of the .4 release of cuckoo to be publicly released shortly I figured I should get this post out as some of the things I talk about here are said to be addressed and included in that release.  If you don't want to wait for that release or something I touch on here isn't included in that release then hopefully the information below will be of use to you. In full disclosure, I'm not a python guru so if you see something that could have been done an easier way or something turns out not to be working for you please let me know...I found out the hard way python is strict on spacing.  Throughout my testing it all seemed to work fine for me but there may be some scenario I didn't test or think of.

(patches available on my github)

General Notes

The installation notes are pretty straightforward to get you up and running and after you successfully do it the first time, any subsequent installation process should be even faster for you.  There are a couple of notes worth mentioning though:

• The first user you create during your Ubuntu installation is an admin user.  This is important to remember if you want your cuckoo user to be a limited user.
• When you add the cuckoo user to its group, you need to log out and log back in for it to take affect.
• To ensure there are no permission issues, you should do the virtualbox setup as the cuckoo user instead of another admin/root account.
• If during your analysis the VM isn't able to be restored or you need to kill cuckoo.py then you need to run virtualbox after and take the vm our of 'saved' mode by discarding it.
• If you are installing 3rd party applications (and you should be if you want to test exploitation), make sure you're properly pointed to them within their appropriate analyzer file "/path/to/cuckoo/setup/packages"
• There's a default list of hashes for common programs that are automatically discarded in the dropped files section so be aware of them "/path/to/cuckoo/shares/setup/conf/analyzer.conf"

Patching

Instead of re-posting all of the files in the cuckoo repo I decided the easiest way to go about releasing these patches/modifications was to utilize the diff & patch commands in *nix. To create the patches:

diff -u 'original' 'new' > 'file.patch'

and once the patches are downloaded from my github, all you need to do is run:

patch '/path/to/original/cuckoo/file' < 'file.patch'

Customization

Web Reports/Portal

At first I couldn't understand why I was able to continuously reanalyze a sample but when I thought about it , it made sense.  Since cuckoo gives you the ability to analyze a file in multiple VM’s, it has to be processed more than once (duh)…maybe a better approach would be to only have that sample be analyzed once by the same VM.

In the main web portal page you are presented with a single search box to search for a files MD5 hash. For convenience and as a time saver I hyper-linked the files MD5 hash in the general information section as well as the dropped files section so you can quickly see if/when it was analyzed previously instead of having to copy and paste it in the main search box every time.

I didn’t want to clutter up the general information section of the report with all of the scans and lookups I was adding to the report so I created two other sections for the report (signatures & lookups).

Signatures

Within the signatures section I added the following ClamAV (2 versions) and YARA.  If you have other scan engines you wish to run against your files then the same type of method could be re-used.  With all three of these features you need to configure the location to their corresponding signatures within "/path/to/cuckoo/processing/file.py".

ClamAV

(besides for above noted change, you also need to edit the path to your clamscan)
I’m a fan of ClamAV and the numerous ways it can be leveraged just make it ideal to have included in my automated processes.  If you’ve read the Malware Analysts Cookbook  (MACB) you might recall that there’s some really handy code made available and one of which shows how to do exactly what I wanted to do – scan the files with ClamAV and show the results.  I don’t like to re-do what someone else has done if it works how I need it to so I made one or two modifications and plugged it in as necessary.

Custom ClamAV

Using the traditional signatures database from ClamAV is good but it can also be worthwhile to create some of your own signatures (remember how logical signatures can be a big help) so I also added a section where you can point it to your custom ClamAV database so it can pickup on other signatures you’ve personally written/acquired.

YARA

On the cuckoo mailing list I came across another user who said he had patches for implementing YARA into cuckoo.  If you’ve read any of my past posts or follow me on twitter you’ll know that I’m a fan of YARA’s capabilities and as such contacted him to see what he had wrote.  The patches themselves were very straightforward and since they worked I didn’t see a need to change them.  He provided me a link to them on his personal GDrive so if you only want to implement that feature into cuckoo then you can use his files, however, the files I’m releasing have that already implemented so no need to do double the work otherwise.  When/if more than one YARA rule is matched, they'll be comma separated within brackets.  The additional files needed besides for for the ones in my github that you'll need to download and install are:

• http://yara-project.googlecode.com/files/yara-1.6.tar.gz
• http://yara-project.googlecode.com/files/yara-python-1.6.tar.gz

Lookups

The looksups section only contains two actual lookups at the moment but also contains what I refer to as ‘future linkage’.  I didn't add the lookups section to the dropped files section because I plan on analyzing them automatically with the modifications mentioned earlier and that would just be too repetitive and a waste of a time.  As far as actual lookups I put in Cymru and VirusTotal for right now so if there’s Internet they will pull the last time the sample was scanned/seen with their services and the A/V detection rate (note - I'm only querying for the hashes, I don’t like submitting for a few reasons).

Team Cymru 

Team Cymru offers a couple of very useful services and one of which I use during investigations is their Malware Hash Registry (MHR). MHR will take the hash(es) you supply it and tells you if it’s a known bad file, the last time they've seen it and an approximate percentage for A/V detection. MACB also had a recipe for adding this to a script so once again I just modified as necessary and inserted to fit cuckoo.

VirusTotal

There are a few scripts online to utilize VirusTotals API and submit/query their site but I decided to use this script.  You can use any method you'd like but if you use the patches I provided just install that script and supply your API key in "/path/to/cuckoo/processing/file.py".  I didn’t want to overly insert code into the existing cuckoo files so I opted to build this file and then import it from within cuckoo.  Essentially I take the files hash and try to get a report of it and if it exists just pull last scan date and detection rate.  While it can be useful to see what the A/V's detected it as, I didn't want to waste time making a collapsable table including all of this information if the new release of cuckoo will already do this.  If it doesn't, then I'll re-visit it.

If the sample doesn't have any VT detection or exist then I have it just state that and if there’s no current Internet connection then state an error.  The latter is very important because I’ve seen others trying to stuff this capability into their code but they fail to address the scenario when there’s no Internet connectivity and therefore their report will fail to be created because they don’t handle the error created.  I wrote it so it would be generic in catching an error because I don’t want my report to fail because of this so if there’s no Internet connection or another error (note that this will also suppress the error that your API key may be wrong!) and the rest of the report is fine to generate then it can still generate.  The same hold true for the snippet for the Cymru check.

Internet connection and results found :

No Internet connection :

Internet connection and no results found :

Future Linkage

I thought it was useful to pre-link the samples to common online sites people use for additional reference/analysis (malwr, shadowserver and threatexpert).  Instead of slowing up the analysis by trying to pull down all of these reports if they exists then parse them I decided it was just easier to create a link for them based on the samples hash that way even if the sample hasn’t been analyzed on any of these sites at the time of my analysis, I could go back to them at a later time and check if a report exists since then.  Just another way to save some time and make life easier.

Dropped Files

Cuckoo will take any dropped files during the analysis of the sample and copy them back over to the host machine under the structure "/path/to/cuckoo/analysis/<#>/files".  By default those files are just left in that subfolder and not analyzed (they will have basic information such as file type and hash in the report though) but I felt it didn't make sense to just leave them in that sub-directory (at least for my goals) so I added the following opted to change "/path/to/cuckoo/processing/data.py" so it would take those files and move it to my samples directory (/opt/samples):

 shutil.move(cur_path,'/opt/samples')

This samples folder is the folder that I'm going to monitoring for new/created files and automatically process them to be analyzed as mentioned later via the watcher.rb script.  Once I did that I noticed another side affect... if there was a queue in the samples directory and the files being moved from the dropped files folder to the samples folder were the same then it would crap out.  I thought the move command would overwrite it but it didn't.  I figured this could be fixed by either copying instead or what I chose to do, check if it exists and if so just delete it from the dropped files folder since it was going to be processed anyway:

        check = os.path.join('/opt/samples/', cur_file)
        if os.path.exists(check):
            os.remove(cur_path)
        else:
            shutil.copy(cur_path,'/opt/samples')
            os.remove(cur_path)

        return dropped

This may not be something that everyone feels they want to do since one obvious consequent I could think of was that since every file is being moved out of the dropped files directory, any special configuration file etc. that you might be interested in won't be there (unless you do file type identification and only move files which can even be processed or if a file can't be processed, move it out of the samples directory to another folder to store dropped files that couldn't get processed i.e. - html files, js etc.).  Another reason might be because it may end up being a continual loop.  Some malware will go out and download another copy of itself etc. and as such by continuing to automatically analyze them will just cause a loop.  This will vary of course by sample, if the Internet is connected and what you want out of your analysis.  Other than that, your analysis task numbers might rise quickly but that shouldn't be on concern because you aren't going to have a sequential set since there's going to be times when a file can't be processes.

Samples Directory Watcher

Melissa wrote a post a little bit ago on integrating cuckoo with NTR and in that post she touched upon the usefulness of having a script running to automatically realize that a new file was created or moved to a certain directory and then take action on that file.  I thought it was nifty and since it was already built into Ruby, I wasn't going to try and hack something else together and see how it held up.   I've read that INotify can be a memory hog so that's something that should be paid attention to although I haven't had any noticeable issues thus far.  If you read the original post you'll soon realize there's some typos... Melissa pointed one out but there are a couple others that might make you frustrated when troubleshooting and to make things easier, I took care of them already.  To get this directory watcher up and running do the following:

sudo apt-get install ruby rubygems 
sudo gem install rb-inotify  

Update 07-18-12 : ... had the wrong command for installing rb-inotify

Download the modified watcher.rb script (on my github too) and edit it to point to the directory you want to watch and the script you want it to execute upon an action/event occurring.  Instead of having an interim script here you can just pass the new sample to "/path/to/cuckoo/submit.py" but I realized I needed an interim script because the sample might be password protected or in a format that cuckoo wouldn't take (i.e. an archive file).

That's the basic customization you need to do for this script, however, you can change it as you see fit.  Initially when I was talking to some Ruby gurus they said that using the IO.popen method was overhaul for what I wanted to do since all I'm essentially doing is passing along a string (new file created/moved) to another file to process.  For testing purposes, I changed it to use exec instead... which worked, but would kill the watcher script after each event.... and that basically killed the purpose of me even having it running so I opted to keep the original method. Once you have all of the pre-reqs installed and the script modified to your needs just open another tab in your shell and let it fly (you don't need the '&' at the end but I like to get my terminal back):

ruby watcher.rb

Archive Parser

If you’re like me then you might have some emails which contain malware samples as attachments or download/get sent password protected archives with possible malware.  If you hand cuckoo an archive or email file (pst etc.) then nothing will happen as it doesn't have a default module to handle them.  As far as the email situation goes, the sheer thought of individually saving each sample one by one doesn’t sound like fun so figured within the interim script I'm calling from the watcher script that there would be a check for a Microsoft Outlook data file and if so, run pffexport against the file.  The thought process is to basically just recursively extract everything out of the the email messages and attempt to process them with cuckoo (if you install libpff, remember to sudo ldconfig after you install it).


To address the archives/pw protected archives issue I try to identify it as an archive file and if so, try to unzip it both with and without a password.  I wasn’t aware that if you supply a wrong password to unzip a file with 7zip that it will still unzip the archive if it turns out that there isn’t even a password protecting the archive (thanks Pär).  I also have a little array set up which contains some of the common password schemes used to password protected malware archives that way I could also add to it in the future (sort of like a dictionary).  

 Additional Software

Depending on the installation you're performing and what additional features you're going to be installing there might be some additional software required which could include:

• YARA
• sudo apt-get install libpcre3 libpcre3-dev 
• python
• sudo apt-get install python python2.7-dev python-magic python-dpkt python-mako
• ssdeep
• http://sourceforge.net/projects/ssdeep/files/ssdeep-2.8/ssdeep-2.8.tar.gz/download
• svn checkout http://pyssdeep.googlecode.com/svn/trunk/ pyssdeep
• g++
• sudo apt-get install g++
• subversion
• sudo apt-get install subversion
• 7zip
• sudo apt-get install p7zip

To-do/Wish List

• The cuckoo DB that's created "/path/to/cuckoo/db/cuckoo.db" only stores a limited amount of information within it.  Even though information regarding a files SHA1/256 hash, ssdeep hash, mutexes, IP/domains etc. are included in the samples report, they aren't stored in the DB.  This helps keep the DB to a limited size but doesn't help if I want to search my repository of analyzed samples for all samples which called a particular IP/host etc.  I didn't want to start changing big chunks of the code to implement this at this point because updates may kill it etc... so I think the better solution will be to only change the snippet which says which fields to create in the DB and to store other selected fields into that DB after analysis.  Another solution can be used to query that DB as it's a common task many of us do anyway.
• The file identification process for determine what type of file the sample is and if it should be processed is pretty basic at this point.  It does the job but at times could use a boost.  A similar thing noticed is if there's certain characters in the samples file name then it won't get processed.  This looks like it could be a one or two line fix with something like Python's string.printable .
• After talking with one of my friends about cuckoo he noted that he's observed not all of the dropped files from the sample being analyzed were being copied back over to the host after the analysis.  This is no bueno... and while I haven't verified this at this time, a simple solution looks to be installing CaptureBAT on the Windows VM and using something (xcopy or robocopy) to copy all of the files caught by CaptureBAT back over to the host after analysis.
• I'm debating to add a switch so I can choose for the analysis to either run wild on the Internet or feed it something like INetSim for simulation.  There are pros and cons to each scenario and maybe a better solution is to use something like Tor ... but I'm up in the air.  As a side note, installing INetSim can be a pain and I'm spoiled as I'm used to it already being installed so other options to look at could be something like HoneyD. 
• I'd like to modify some of the existing analyzers to run additional programs against a sample and report on their results (i.e. hachoir-subfile, pdfextract etc.)

XDP files and ClamAV

This has been ported over to my GitHub site and is not longer being maintained here. For any issues, comments or updates head here.

updated 08/20/2012 - added two new signatures

There were some recent discussions going on regarding the use, or possible use of bypassing security products or even the end user by having a XML Data Package (XDP) file with a PDF file.  If you aren't familiar with XDP files, don't feel bad... neither was I.  According to the information Adobe provides, this is essentially a wrapper for PDF files so they can be treated as XML files.  If you want to know more about this file then take a look at the link above as I'm not going to go heavily into detail but note that the documentation is a bit on the light side as it is.  There're other things that can be included in the XDP file but for this post we're looking at the ability to have a PDF within it.

 Adobe states that :

"The PDF packet encloses the remainder of the PDF document that resulted from extracting any subassemblies into the XDP.  XML is a text format, and is not designed to host binary content. PDF files are binary and therefore must be encoded into a text format before they can be enclosed within an XML format such as XDP. The most common method for encoding binary resources into a text format, and the method used by the PDF packet, is base64 encoding [RFC2045]."

Based on my limited testing, when you open a XDP file, Adobe Reader recognizes it and is the default handler.  When the file is opened, Adobe Reader decodes the base64 stream (the PDF within it), saves it to the %temp% directory and then opens it.

Brandon's post included a SNORT signature for this type of file but I wanted to get some identification/classification for more of a host based analysis.  Since I couldn't get a hold of a big data set I grabbed a few samples (Google dork = ext:xdp) and thought I'd first try TrID - but that generally just classified them as XML files (with a few exceptions) and the same thing with 'file'.  I can't blame them, I mean they are XML files but I wanted to show them as XDP files with PDF's if that was the case - that way I could do post-processing and extract the base64 encoded PDF from within the XDP file and then process it as a standard PDF file in an automated fashion.  

I then looked to TrIDScan but unfortunately that didn't work as hoped.  I tried creating my own XML signature for it as well but kept receiving seg. faults .. so... no bueno. My next thought was to put it into a YARA rule but I thought I'd try something else that was on my mind.  I've been told in the past to mess around with ClamAV's sectional MD5 hashing but that's generally done by extracting the PE files sections then hashing those.  Since this is a XML that wasn't going to work.  I remembered some slides I looked at a bit ago regarding writing ClamAv signatures so when I revisited them the lightbulb about the ability to create Logical Signatures came back to me.

ClamAV's Logical Signatures

Logical Signatures in ClamAV are very similar to the thought/flow of YARA signatures in that they allow you to create detection based on..well.. logic.  The following is the structure, the 'Subsig*' are HEX values... so you can either use an online/local resource to convert your ASCII to HEX or you can leverage ClamAV's sigtool (remember to delete trailing 0a though):

 sigtool --hex-dump

Logical Signature Structure:

SignatureName;TargetDescriptionBlock;LogicalExpression;Subsig0;Subsig1;Subsig2;...

 

Looking back to Adobe's information they also mention that the PDF packet has the following format:

<pdf xmlns="http://ns.adobe.com/xdp/pdf/">
     <document>
          <chunk>
               ...base64 encoded PDF content...
          </chunk>
     </document>
</pdf>

ClamAV Signature

The beauty is that you can create your own custom Logical Database (.ldb) and pop it into your default ClamAV directory (i.e. /var/lib/clamav) with the other databases and it'll automatically be included in your scan. While just detecting this may not indicate it's malicious, at least it's a way to detect the presence of the file for further analysis/post-processing.  So based on everything I now know I can create the following ClamAV signature :

XDP_embedded_PDF;Target:0;(0&1&2);3c70646620786d6c6e733d;3c6368756e6b3e4a564245526930;3c2f7064663e

Explained: 

XDP_embedded_PDF - Signature name

Target:0 - Any file

(0&1&2) - match all of the following

0
ASCII :  <pdf xmlns=
HEX  : 3c70646620786d6c6e733d

1
ASCII :  <chunk>JVBERi0
HEX :  3c6368756e6b3e4a564245526930

* JVBERi0 is the Base64 encoded ASCII text " %PDF- ", which signifies the PDF header.  It was converted into HEX and added to the end of the 'chunk' to help catch the PDF

2

ASCII :  </pdf>

HEX :  3c2f7064663e

update #1 on 08/20/2012 : 
The above first created ClamAV signatures works but I started to think that the '<chunk>JVBERi0' may not be next to each other in all cases ... not sure if they have to nor not by specification but this is Adobe so I'd rather separate them and match on both anyway..

XDP_embedded_PDF_v2;Target:0;(0&1&2&3);3c70646620786d6c6e733d;3c6368756e6b3e;4a564245526930;3c2f7064663e 

update #2 on 08/20/2012:

YARA signature:

rule XDP_embedded_PDF
{
meta:
author = "Glenn Edwards (@hiddenillusion)"
version = "0.1"
ref = "http://blog.9bplus.com/av-bypass-for-malicious-pdfs-using-xdp"

strings:
$s1 = "<pdf xmlns="
$s2 = "<chunk>"
$s3 = "</pdf>"
$header0 = "%PDF"
$header1 = "JVBERi0"

condition:
all of ($s*) and 1 of ($header*)
}


Questions to answer

Actors are always trying to find new ways to exploit/take advantage of users/applications so it's good that this was brought to attention as we can now be aware and look for it.  While the above signature will trigger on an XDP file with a PDF (from what I had to test on), there're still questions to be answered and without having more samples or information they stand unanswered at this point:

1. Could these values within the XDP file be encoded and still recognized like other PDF specs
2. Can it be encoded with something other than base64 and still work
3. Will any other PDF readers like FoxIT treat them/work the same as Adobe Reader

Comments and questions are always welcome ... never know if someone else has a better way or something I said doesn't work.

Deobfuscating JavaScript with Malzilla

This has been ported over to my GitHub site and is not longer being maintained here. For any issues, comments or updates head here.

I was asked a question a little while ago from a fellow forensicator about deobfuscating some JS that he came across.  The JS didn't take long to reverse but I suspect there are others out there that would benefit from a quick post regarding another way to go about this task.  While there's jsunpack, js-beautify etc. I chose to run it through Malzilla for this example.

The structure of the JS was noticeably familiar and turns out to be related to an exploit pack; which is a common source of where a lot of the JS you might come across in the DFIR field results from these days.  These types of kits make it point-and-click easy to not only distribute malware but also make it uber-easy to obfuscate the code on their pages.

The first thing to do is copy out what’s in between the ‘<script>’ tags and place it in the top box of the ‘Decoder’ Tab within Malzilla - we don't need the other <html> tags etc., we only need the goods.  Next step is to get rid of what we don’t necessarily need at this point (shown commented out with ‘//’).  This will vary depending on what you're analyzing and may take a bit more knowledge to realize but just remember what your goals are - there will be junk thrown into the mix and since all I care about at this point is to see what gets produced (URL etc.) the top part didn't look relevant for helping me get my question answered :

At this point you have a few options (1) replace the eval() (2) run it through debugging to verify it's working (3) run the script.  Everything looks good enough to work so let's just go ahead and choose to run the script:

Note that even though the bottom text displays “Script can’t be compiled” (seen above) … the eval results were still produced.  To see the results, click on ‘Show eval() results’ then double click on each of the results (one in this instance) and the results will be displayed in the lower pane – this time showing the produced iframe :

There's generally always more than one way to get the results you require so hopefully this will help some of you next time.