Announcing 𝟘 𝟙 ❒ 🔨 Strengthen cryptography with AI. Cryptography—both current and post-quantum cryptography—depends on hashes. SHA-256? SHA-3? BLAKE3? Nobody knows if they are secure. We're going to find out! Come help accelerate scientific research! HashSmash is an open science-acceleration project in collaboration with Yukon, Eigen Labs, Project Eleven,... and Shielded Labs.show more

great to work with you on one of the hardest problems in the world and to accelerate scientific research, zooko! thank you for all the support ❤️

Coordination to solve a tough scientific problem begins here! We’re so excited for this opportunity.

Let’s accelerate one of the HARDEST problems to exist!

very cool! congrats on the launch 😁

Super cool

nice I predict sha256 will fall >:) sha3 will remain

God wills it

veryyy cool

great initiative!

Do you control your AI agents or do you trust them without vetting them?

The tracks target SHA-256 at 31-32 of 64 rounds, SHA3-256 at 5-6 of 24 and BLAKE3 at 1-2 of 7. All five SHA-256 entries in review cost 2^128 or more, so none beats the generic birthday bound yet.

strengthen cryptography with AI, now the bugs write themselves

Great stuff! We need more such automated research efforts in the age of AI.

🔥

Hashes are the one part of crypto that survives Shor intact. Breaking SHA-256 or SHA-3 would be the first AI cryptanalysis result that actually matters. This is the right fight.

What a time to be in cryptography. AI agents stress-testing our hashes in public?

superb!

I'm more and more convinced that future direction of research is autoresearch.

huge!

Amazing! Shielding the future.

Worth separating the two things this can produce, because they are not symmetric. A break on a reduced-round target is real information. It verifies in microseconds, it needs no judge, and it stands on its own forever. A non-break is almost no information, because the search was bounded: by compute budget, by the window, and by the strategies the models were steered toward. The non-break is also the result that will travel, as "AI went at SHA-256 and could not move it." Confidence in a hash has always been assembled out of accumulated failed attacks, and that works because the attackers are independent and keep coming back for decades. Failures inside one tournament are correlated with each other and with the tooling, so they buy much less confidence per unit of failure than the usual intuition assumes. Which makes the most valuable artifact here the search log rather than the leaderboard: compute charged per target, strategies tried and abandoned, where the dead ends were. A null result is only interpretable if the shape of the search is published next to it.

@lucidzk Smashed the like button

Nice🔥

Test hashes before whining 👀

Finally putting AI to work on actual security.

AI will change

Open science for crypto breakage is a genuinely great idea, GPUs are about to get very good at finding collisions.

Hashes carry a huge amount of the weight in post-quantum designs, and they get far less public stress testing than the signatures do. Pointing automated search at reduced round versions sounds like the right place to begin, since that is usually where the first cracks turn up 👽

congrats on the launch, make cryptography invulnerable

BLAKE3 in the crosshairs is spicy

Crypto Agility