The model found a previously unexploited symmetry in the lattice HAWK relies on, and restoring the intended security level would mean doubling key sizes — no deployed system is affected.
Sixty hours of work by the Claude Mythos Preview model improved on the strongest attack previously known against HAWK, a signature scheme built to withstand quantum computers, and in effect halved the strength of its keys. HAWK is among the designs that reached the third round of NIST's call for further signature schemes, and it had come through two rounds of scrutiny by human experts over two years before the model was pointed at it. Anthropic's announcement, made public on July 28, sets out two results from that model: one against a version of AES with a reduced number of rounds, and the attack on HAWK.
That attack proceeds by locating a nontrivial automorphism — a symmetry in the lattice the scheme depends on, one nobody had taken advantage of before. The symmetry allows a quicker enumeration attack, exponential still, and it means HAWK's key sizes would have to double to hold the same security level, which strips away much of what made the scheme appealing as a post-quantum candidate. At the smaller HAWK-256 size, recovering a key in full had been assumed to cost far more than Mythos showed was actually necessary; against HAWK's larger key sizes, this kind of attack stays impractical.
The other result concerns AES, the symmetric cipher NIST took up as a standard in 2001. Few encryption algorithms have drawn closer examination; researchers routinely probe deliberately weakened variants of it to gauge how sturdy the complete algorithm is. Mythos worked out how to break one of those weakened variants by removing one of the guesses an attacker would otherwise have to make, running 200 to 800 times faster than the best attacks known before.
Neither finding matters in practice for systems in use today: HAWK is still only a proposal and has not been fielded, and the AES attack applies to a cut-down form of the cipher, leaving the full standard unbroken. The company characterizes both as 'expected' results — NIST's standardization process exists precisely to surface such weaknesses in candidates before deployment, and the AES work extends a long line of prior research against reduced-round variants.
What has changed is the level at which the flaws sit. At its initial release, the model could already find and exploit security holes without human help in nearly every piece of software it was tested against, several major cryptographic libraries among them — but those holes came from how the code was written, not from defects in the algorithms behind it. These results are flaws in the mathematics of the algorithms themselves. Claude reached both largely on its own: one of the company's researchers worked with it to build the HAWK attack, while a colleague constructed a scaffold through which Claude turned up the AES attack entirely by itself, at about $100,000 in API spending per result.
The constraint has moved to checking the work. Anthropic says the bulk of its effort over recent months has gone into confirming the results are correct. The HAWK attack is simpler to check because it runs from start to finish; two researchers needed close to a month before they were sure the AES method Mythos found in a single week was sound. The company expects academic cryptography to come under the same pressure cybersecurity already faces, where models surface more software bugs than the customary human routines for sorting, confirming and fixing them can absorb.
These are the most powerful attacks the company says it has produced so far, and it made them public only after a stretch of consultation with leaders in the US government and in industry. Its argument is that there is good reason to work out now what researchers ought to do if a language model turns up weaknesses in cryptosystems that bite in the real world straight away — a question it says academia, government and industry will all need to weigh in on. It said it would run a workshop for academics within the next few weeks on what language models mean for research in security and cryptography.