Tuesday, August 21, 2018

Crack me if you can 2018 write-up

Crack me if you can write-up 2018



Active participating members
15
GPUs equivalent to GTX1080 peak
60
GPUs equivalent to GTX1080 constant
40
CPU threads peak
1300
CPU threads constant
600
Contest related Instant Messages sent
~7000
Hash:plain submissions to internal platform
>5300
Hash:plain submissions to Korelogic
2293



Members

blazer cvsi espira gearjunkie hops m33x mastercracker milzo jimbas mexx666666 s3in!c usasoft user vetronexe winxp5421





Prep

After hearing news that Korelogic would be awarding bonus points for first unique founds, we took precautions to tune our submission process to ensure we could capitalise on this bonus. To avoid false spam triggers, an alternate email provider that supported bulk inbound/outbound requests was used. In addition, various functions on our hash management platform were disabled and tweaked such that the hash:plain pairs could be processed and uploaded quickly at a constant but not too aggressive rate.  We only had a handful of submission troubles which were rectified quickly on our end.





Patterns

It was quite cheeky for Korelogic to use usernames from the competing teams as plaintexts and this was spotted quite early on in our MD5 list. Similarly, they were seen in the SSHA, MD5(unix) lists, we also noticed that each algorithm was assigned a specific range of starting characters. Seeing as that the other teams were getting bcrypts it appeared that these were possible, and this was where all the points were at.  While some of our members continued to collect points by exploiting the 4x first unique found bonus for the lower scoring hashes, others worked on trying to get a break on bcrypt hashes using the patterns we spotted. It was not long before we found the starting characters for the bcrypt hashes using the usernames in double combo mode.


Strategy
Once we had the first bcrypt hit, we tried to uncover the complete list of usernames from the plains found in the faster algorithms. After we were confident we had a solid pattern, we brought up many CPU crackers running MDXfind to work solely on bcrypt hashes. It was a little chaotic initially as we tried to figure out the best way to distribute the workload for bcrypt hashes. One of our members then stepped up and became the central point for distributing the tasks but the task distribution and request was still done manually. Soon another member whipped up a semi-automated procedure where each member could request custom tasks from a central distribution list. During our peak we utilised roughly 1300 CPU threads but we had around 600 sustained threads throughout the contest. A small cluster of 16 odroids (XU4) running MDXfind-ARM were also used to attack the bcrypt hashes. Sidenote, it was relatively cheap and efficient to attack bcrypts using ARM cores. Each odroid gave us roughly 50H/s (800H/s in total) for the contest’s bcrypt hashes (cost factor 10) and the cluster in total uses approximately 200W. This results in a efficiency of 4H/s/W.

Due to the unfriendly nature of bcrypt on GPU, all GPU resources were reserved for the other 3 algorithms which worked much more efficiently with hashcat GPU. Members were free to decide whether they wanted work on patterns alone which some opted to and devised their own methods and scripts which they used to attack patterns on the algorithms, while other joined the hashtopolis instance which had around the equivalent of 60 GTX1080s.
We were generally quite close score wise with team hashcat and trailed them for the first 15 hours or so into the contest. When one of our members woke up and submitted over 100 unique bcrypts we leapfrogged over hashcat into first place and took a comfortable commanding lead. This was a great morale boost and more CPU instances were placed onto bcrypt as we realized other teams were using different patterns from us and we had identified a very efficient one which yield many hits for little work. Additional patterns were later identified, such as one where popular suffixes (pass01, pass02 etc) were used across all of the algos); though these did not seem as efficient as the username combos.

Some stats from our hash management platform showing rate of uploads

MD5(Unix)


SSHA
MD5


 Bcrypt

After thoughts
We do regret not switching over to JTR for a nice bcrypt speedup when more candidates than cores are used due to its bitslice interleaved implementation, yielding up to twice the speed over MDXfind. We also failed to spot the full range of starting characters for bcrypt and lost some valuable points there too.

Towards the end we tried to spread the attacks across all the algorithms so we would not only be ranked highest by score but also highest across algorithms. This was quite hard to maintain as it seemed like both team hashcat and john were gaining ground on us. Overall, we were quite impressed with our ability to obtain more unique bcrypt firsts than both john-users and hashcat combined which allowed us to take first place. A massive thanks to Korelogic for hosting the contest once again, we really enjoyed the added twist this year as it gave us all an incentive to constantly submit. A shout out to our competitive rivals, Team Hashcat and john-users for pushing us hard and making us drink that extra cup of coffee to stay up.

Looking ahead
We have enjoyed playing CMIYC over the years. So, when presented with the opportunity to create our own password cracking contest we jumped at the idea. In 2019, we will be hosting our own CMIYC style contest at Cyphercon in Milwaukee, WI. We hope all of you will join us for the first “Crackthecon”. As more information about the contest is finalized we will update the contest site crackthecon.com.



No comments:

Post a Comment