Here is a 72-byte alphanum MD5 collision with 1-byte difference for fun:
md5("TEXTCOLLBYfGiJUETHQ4hAcKSMd5zYpgqf1YRDhkmxHkhPWptrkoyz28wnI9V0aHeAuaKnak")
=
md5("TEXTCOLLBYfGiJUETHQ4hEcKSMd5zYpgqf1YRDhkmxHkhPWptrkoyz28wnI9V0aHeAuaKnak")
I once used the cluster of 215 PlayStation 3s at EPFL, that was a lot of cheap computing power at the time. And unlike for other CPUs, assembly programming for SPUs was quite magical: exact clockcycle predictions just from code.
This is the first md5 collision with only printable ascii that I know of.
I have been asked before if this was possible, but I used to respond its not practically doable.
I'm very proud and thankful to have won one of the RWC2020 Levchin prize together with Xiaoyun Wang for our work on hash function cryptanalysis!! #realworldcrypto
Its an identical-prefix collision attack where you can pick your own allowed charset (say alphanum, base64, all printable). It also allows to force some specific bytes (mainly 0-7 and 20-27) to some extent.
Seriously, stop using SHA-1! SHA-1 chosen-prefix collisions are now practically demonstrated. Beware of ALL possible collision exploits. E.g. see the amazing list of PoCs by @angealbertini.