Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Date: Tue, 30 Oct 2012 19:10:15 +0100
From: magnum <john.magnum@...hmail.com>
To: "john-dev@...ts.openwall.com" <john-dev@...ts.openwall.com>
Subject: Serious bug in md5crypt-cuda - or in driver?

The current md5crypt-cuda code works fine with Fermi and Kepler cards but on a 9600GT I noticed today I got unreasonably good speed:

Benchmarking: md5crypt [CUDA]... DONE
Raw:	63351K c/s real, 63351K c/s virtual

At first I did not spot anything unusual but hey wait a minute, is that a K tucked on the figures?!

So I ran it through the Test Suite which very quickly ends up claiming everything passed. But some manual testing revels that it will simply accept ANY password as valid for ANY hash. We should enhance the Test Suite (and possibly the self tests?) so it can detect false positives like these.

This is not necessarily a bug in Lukas' code, it may be in whatever version of CUDA I'm running on that machine so I'll need to investigate more. But anyone having pre-fermi cards, please share your benchmarks or real tests. BTW all other CUDA formats seem to work fine on that card except mscash, wpapsk and pwsafe - these three fails self-tests. All three pass on my laptop GT650.

Oh, and while looking at the code I saw an unrelated problem in set_salt(): It resets the global any_cracked variable without clearing the outbuffer - and this causes such clearing to be skipped in the next crypt_all(). In rare cases this will probably also lead to false positives but this is not the cause of the initial problem.

magnum

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.