Skip to main content

Welcome to Mark Baggett - In Depth Defense

I am the course Author of SANS SEC573 Automating Information Security with Python. Check back frequently for updated tools and articles related to course material.

TCP Fragment Evasion

Originally posted on

By: Mark Baggett

I recently read a very good article on tuning Snort's Stream5 preprocessor to avoid "TCP Fragment Overlap" attacks. It's a great article, but the wording confused me. I thought to myself, "TCP Fragments, that must be a mistake. The TCP Header doesn't have a 'more fragments bit', a 'fragment offset' or anything to support fragmentation. How can there be any TCP fragments?"   Typically when we talk about fragmentation attacks we think about Layer 3 attacks. Attackers manipulate the IP packet headers to pull off various insertion and evasion attacks. Examples of layer3 attacks include overlapping fragment attacks and temporal evasion (host reassembly timeout evasion). These attacks are explained pretty well in an article titled "Evading NIDS, revisited".

So what is TCP or Layer 4 "fragmentation"? Really, its overlapping or retransmitted datagrams with the same TCP Sequence number. To demonstrate the concept I fired up a virtual machine running Backtrack 3. I ran a netcat listener on my host (nc -l -p 9000) and used a netcat client in backtrack to connect to it. I fired up wireshark to watch the packets and I transfered the text "This is a test of the emergency broadcast system. If it were an actual emergency" between the two hosts. This is what Wireshark captured.

Figure #1


View image

Perfect. Exactly what we would expect. Since my packet doesn't exceed the MTU of the established TCP connection a single packet is transfered to the client with a single acknowledgment in return.  If it had exceeded the MTU it still wouldn't have fragmented. It would have sent more than one datagram, each with its own unique IP ID.

Then I created a fragroute configuration file with one line in it:

tcp_seg 16

This will cause fragroute to break the packets down so that they can only carry 16 bytes of TCP traffic. I start fragroute (fragroute -f ~/myfrag.conf and transfer the same text between the hosts...

Figure #2


View image

Fragroute works as expected and breaks the packets down such that only 16 bits of data can be transfered in each packet. Each packet sequence number increases by the number of bytes transmitted. Sequence numbers increase in order. Also, notice that each packet has its own unique IP ID field. There is NO FRAGMENTATION. The "More Fragments bit" isn't set.   The fragment offset isn't set. No fragments. Instead, fragroute is transferring packets as if the MTU of the segment is only enough for 16 TCP bytes.

So now lets do some "tcp fragmentation overlaps". I change my fragroute.conf file to say this:

tcp_seg 16 new

This will cause fragroute to transmit frames with overlapping sequence numbers.   This attack takes advantage of the fact that the TCP layer doesn't pass data up the stack to the application until it has acknowledged the data and that packets are acknowledged in sequential order.    So if we skip datagram #3  and transmit datagrams #4, #5 and #6, duplicates of #4 and overlaps of #5 and #6 then the TCP stack needs to hold datagrams #4,#5 and #6 (as long as they are within the window size) and figure out what to do with duplicates/overlaps once it receives fragment #3.

To see this in action I fire up fragroute and retransmit the text "This is a test of the emergency broadcast system. If it were an actual emergency"  

Figure #3


View image

Lets look at it in the fragroute packets in figure#3.   The first two datagrams (#1 and #2) are garbage.  Their payload is random junk.  Then fragroute transmits good data in packets 4 and 5.  The payload here is the end of our payload "If this had been an actual emergency."  After the 4th packet the receiving host begins screaming to the transmitting client "HEY DUDE,  ACK 2933750986.  I didn't get that one yet".   The receiving TCP stack is complaining about not receiving the first datagram.    Then fragroute sends 2 packets with 32 TCP bytes in each.  These two datagrams include the  FIRST datagram (Notice packet #10 has the lowest sequence number and the embedded text payload).  Parts of these two packets overlap packets 1 and 2. Packet #9 overlaps 16 bytes of packet #2.  16 bytes of packet #10 overlap packet #1.  If the TCP reassembly engine favors NEW packets then it will reassemble the text as expected.  If the IDS reassembles the packets favoring the OLD packets then we can bypass the IPS. If we were drawing analogies to layer three fragment attacks holding the low sequence number datagrams is equivalent to setting the "more fragments bit" and the sequence number is the equivalent to the fragment offset. So how to fix this?  The attacks aren't new.  Snort has the STREAM5 preprocessor.  Just be sure that you tune STREAM5 just like your FRAG3 preprocessor.  


Snort's Stream5 and TCP overlapping fragments An article by Richard Bejtlich that sparked my interest in this topic. Its a very good article with more explanation on tuning the snort preprocessor.

Popular posts from this blog

Awesome Keyboard Tricks - Clevo/Sager Backlight control from Powershell

I'm back on Windows.   After 8 years on a Macintosh I just couldn't go another day with ONLY 16GB of RAM.   I priced it out and for the cost of a top of the line MacBook I could get a tricked out PC with 32GB of ram and 2.5 TB or hard drive space (1.5 of it being SSD).   So I made the switch.  To get a top performing laptop I ended up buying a gaming machine from   The model is Sager NP9752 ( Clevo P750ZM ).    I have to say I like it quite a bit.    One of the features I was curious about was the "Programmable backlit keyboard".   With it you can set your keyboard backlight to various colors and light movement patterns.    Now, when I hear "programmable" I think APIs.   I was a little disappointed to find out there weren't any documented APIs that I could use to control the keyboard.    Your only choice is to use their built in tool to configure the lights on the keyboard.   That stinks.  I want to be able to change key colors automatically

SRUM-DUMP and SRUM_DUMP_CSV Ported to Python 3

SRUM_DUMP and SRUM_DUMP_CSV have been ported to Python3 and are available for download from the PYTHON3 branch of my github page. In moving to Python3 I also updated the modules that I depend upon to parse and create XLSX files and access the ESE database that contains the SRUM data.  I hope that this will fix the issue that some users have experienced with SRUDB.dat files that create very large spreadsheets.  If it does not please let me know and continue to use SRUM_DUMP_CSV.EXE to avoid the XLSX problem. In moving to Python3 you will find the process to be faster. If you would like to run the tools from source instructions for doing so are in the README on the github page.

New tool

I read an article on Fireeye's website the other day where they used Machine Learning to eliminate a lot of the noise that comes out of tools like strings.  It's pretty interesting and looks like it would save me some time when looking through malware. I wondered how effective scores would be in helping to eliminate the noise.  45 minutes and 29 lines of Python code later I have something that looks like it works.  Check out Before here is the output of strings on a piece of malware: student@573:~/freq$ strings -n 6 malware.exe | head -n 20 !This program cannot be run in DOS mode. e!Rich `.rdata .pdata @.gfids @.rsrc @.reloc \$0u"H L$ SVWH K SVWH |$ H;_ <bt%<xt!<Zt |$ AVH l$ VWAV L$ SUVWH UVWATAUAVAWH 0A_A^A]A\_^] UVWATAUAVAWH @A_A^A]A\_^] After the useful stings quickly bubble to t

Three Free Python apps to improve your defenses and incident response

I did a SANS Webcast on a couple of tools I developed to use automation to enhance your network defenses and incident response capability.    If you missed it you can check it out here:


An issue was reported where is some conditions SRUM_DUMP would stop processing and print the following error to the screen. UnboundLocalError: local variable 'sid_str' referenced before assignment The issue was that sometimes the SRUM database had entries in it that were all zeros. OrderedDict([('IdType', 3), ('IdIndex', 38127), ('IdBlob', '0000000000000000')]) I've released an update that handles the anomoly althought I do not understand the circomstances of why Windows would record all zero's for as the user SID. The issue was fixed and new versions of both SRUM DUMP and SRUM DUMP CSV were released.