Distributed Hash Cracker

As of the October 23 release, the cracker now supports batched testing (for all unsalted algorithms, i.e. all but MD5crypt) of multiple hashes against a given search space.

Up to 128 hashes may be entered, one per line, in a text box on the controller submission page. The system is fairly efficient; in one two-minute long test quadrupling the number of hashes being tested against resulted in a mere three seconds increase in run time.

I am currently redesigning the network protocol and database schema to allow the cracker to batch-process a large number of hashes simultaneously. More information will be posted as the code approaches completion.

Please be advised that due to a fairly major refactoring of the database schema, the controller will need to be reinstalled (delete config.php, drop all tables in the controller database, and re-run setup.php) to run correctly with the new code.

This release will also break protocol compatibility with current agents. (If an agent connects to a new controller without indicating that it is compatible with the new version 3.2 protocol, it will be told that no work is available.)

While studying the MD4 hash algorithm in an attempt to increase MD4 and NTLM performance, I discovered an interesting vulnerability which will permit cracking of short passwords to run up to 15% faster. The attack has not, as of this writing, been implemented in the cracker.

A brief paper describing the attack can be found at http://www.cs.rpi.edu/~zonena/papers/md4_v2.pdf.

In the off chance thats somebody might find something useful in it, I've decided to post the code that started it all - version 1.0.

This code compiles for Windows only (due to its use of MFC) and is poorly documented. Work units are allocated statically at the start of the crack; if any compute node goes down the entire crack fails.

The source has been gathering dust on my server for months and has not been - nor will it ever be - maintained since the end of 2008. Don't expect it to work properly :)

To prevent confusion and keep the download page tidy, we will no longer be archiving all repository snapshot tarballs. Instead, only the current and previous snapshots will be posted for download.

Anyone interested in older revisions is, of course, free to check out the appropriate files from our Git repository.

The cracker now supports NTLM.

• GeForce 8600M GT - 31.488 MHz
• GeForce 9800 GT - 159.869 MHz
• Tesla C1060 - 347.003 MHz
• GeForce GTX 285 - 434.031 MHz

• Standard test cluster - 2.30 GHz

MD4 support was added to the cracker in record time - 6 minutes and 38 seconds after I sat down at the keyboard, a working CUDA implementation (derived from the MD5 kernel) had been pushed to Git. The algorithm was similar enough to MD5 that it passed unit tests the first time I compiled!

• GeForce 8600M GT - 30.182 MHz
• GeForce 9800 GT - 153.436 MHz
• Tesla C1060 - 339.702 MHz
• GeForce GTX 285 - 423.313 MHz
• Standard test cluster - 2.31 GHz

NTLM, which is an MD4-based algorithm, should follow shortly.

The current version now supports md5crypt on CUDA.

As with the initial SHA-1, performance is truly abysmal (0.09 MHz on a Tesla C1060) due to being memory bound. MD5Crypt is, unfortunately, a more complicated function which will take longer to optimize.

The initial SHA-1 implementation was extremely memory bound, running at only 25 million hashes per second on a Tesla C1060.

After optimizing it now runs at 161 MHashes/sec on a C1060. On my standard test cluster performance hit a truly impressive 969 MHashes/sec.

For those interested in tracking the progress of optimizations, stats from my laptop's GeForce 8600 are below. All speeds are in MHashes/sec.
c95085a5 3.88
9896e693 4.58
9ab76c6f 4.62
ea075fc1 5.44
1876e5f2 13.06

The cracker now supports SHA-1 on CUDA.

Please note that this is an initial implementation and is severely memory bound, running at only 25 MHashes/sec on a Tesla C1060. Further optimizations should bring the speed up to nearly that of MD5.