Anthropic's locked Mythos model just found a real weakness in a NIST post-quantum signature finalist

[Wrench]

Anthropic published research today saying an unreleased version of Claude, internally called Mythos Preview, found two previously unknown attacks on cryptographic schemes, and one of them lands directly on HAWK, a lattice based signature scheme that is currently the last one standing in NIST's third round post quantum signature competition

The actual math is the interesting part, Claude found a symmetry inside HAWK that no human researcher had spotted, and for the smallest key configuration it dropped the cost of recovering a secret key from 2 to the 64 operations down to 2 to the 38, that is roughly 67 million times less work than anyone thought was required

The fix is straightforward on paper, just double the key size, but Anthropic's own writeup admits that doing so wipes out most of what made HAWK attractive in the first place, compact keys and fast signing were its whole pitch against the other post quantum candidates, and blockchains in particular care about signature size because it eats directly into block space and fees

To be clear nobody needs to panic about their coins, HAWK has never actually been deployed anywhere and Bitcoin still runs on ordinary ECDSA, this is a competition candidate getting knocked around before it ever reaches production, which is honestly the best possible time for this kind of flaw to surface

The second result is arguably the more interesting story about how the model actually works, researchers targeted a 7 round research version of AES, blocked it from using all five known families of attack technique, and told it to invent a sixth, the model initially refused outright and said there was nothing easy to find on the most studied block cipher in existence, which is a pretty striking moment of an AI correctly assessing that a problem was genuinely hard rather than just guessing

It only cracked it after researchers pushed back a few times over several days, eventually producing something called a Mobius Bridge that eliminates one of the nine key bytes an attacker previously had to guess, that alone beat a record cryptographers had held since 2013, and it took roughly a billion output tokens and several hundred hours of human verification afterward just to confirm the result was real

Claude Mythos Cracked Post-Quantum Cryptography That Humans Spent Years Failing to Break - Decryptdecrypt.co

[Rob72]

The model refusing at first and saying there's nothing easy to find is the detail that actually impresses me most here, that reads like genuine calibration rather than just an AI bluffing its way through a hard problem

[CarlosBuddle]

Doubling HAWK's key size to patch this basically kills its entire value proposition against the other post quantum finalists, feels like this might knock it out of the NIST race entirely rather than just weaken it

[Darren_20]

To be fair getting knocked out during round three of a standards competition is exactly what this process exists to catch, way better than finding this flaw after a blockchain has already deployed it in production

[Annie]

The LEA result getting almost no attention is a shame, breaking 13 rounds in under an hour on a desktop against a prior best that needed 2 to the 98 plaintext pairs is a massive jump even if the deployed version runs a safer 24 rounds

[Hyperdrive71]

That verification gap is the scariest part of this whole thing to me, a wrong result confidently published as broken cryptography could do real damage if nobody catches it in time

[FluxKnight46]

CryptanalysisBench scoring under 9 percent on ciphers with no published break at all is the number that keeps this in perspective for me, impressive on known hard problems but nowhere near breaking unbroken modern crypto yet

[AJStyles92]

Several hundred hours of human verification for a result the model produced in days is the real bottleneck story buried here, we are heading toward a world where the AI generates cryptanalysis faster than humans can check it is actually sound