On July 28, Anthropic published research showing that Claude Mythos Preview, their unreleased frontier model, found two previously unknown cryptographic attacks in roughly 60 hours of autonomous work. One of them helped take down a NIST post-quantum signature candidate that had survived two rounds of expert human review spanning two years. That is not a prompt engineering trick. That is AI doing original mathematics.
To be clear about what this is and is not: the production systems you depend on are not at immediate risk. Full AES-256 is fine. The attacks target weakened or candidate schemes. But the more important story is how fast this happened, and what it tells us about where AI fits in the research pipeline going forward.
What Claude Mythos Actually Did
Claude Mythos Preview ran for approximately 60 hours on cryptanalysis work targeting two algorithms: HAWK, a lattice-based post-quantum signature scheme that was in NIST's additional standardization process, and a weakened seven-round variant of AES.
For HAWK, the model identified an unexploited mathematical symmetry in the scheme's lattice structure. Exploiting that symmetry improved the best-known key-recovery attack on HAWK, effectively cutting the algorithm's security strength in half. Anthropic disclosed the finding to HAWK's authors in June and coordinated with the NIST mailing list before publishing. On the day of public disclosure, the HAWK team withdrew the scheme from NIST's post-quantum signature standardization process. The community had been reviewing HAWK for two years across two evaluation rounds. The model found this in 2.5 days.
The AES work is more nuanced. This is not an attack on AES as you deploy it. It targets seven-round AES, a reduced-round variant used in academic cryptanalysis to probe the algorithm's structure. Claude developed a technique Anthropic calls the Möbius Bridge: it eliminates one of the guessing steps in prior meet-in-the-middle approaches. Combined with additional model-generated optimizations, the result is an attack 200 to 800 times faster than the previous best on seven-round AES. Full AES uses 10 to 14 rounds depending on key size, so this is not a production threat. But it is meaningful progress on a problem cryptographers have been chipping at for decades.
Autonomy Level: Where It Gets Interesting
The HAWK result is the more remarkable one from an autonomy standpoint. Anthropic says Claude nearly fully independently identified and exploited the lattice symmetry. The AES work involved more human-AI collaboration, with Anthropic researchers guiding the model toward specific sub-problems.
That distinction matters. The model was not just a fast calculator running enumeration. For HAWK it was doing something closer to what a cryptographer actually does: examining a structure, noticing something that prior reviewers missed, then constructing an attack around that observation. Whether you call that "reasoning" or "pattern matching at scale" is a philosophical question. The output is real math that real experts will now have to account for.
What to Do If You're Building on Post-Quantum Crypto
If your stack includes HAWK, pull it. That is the NIST signal.
If you are deploying post-quantum signatures more broadly, the safe ground right now is the schemes NIST has already fully standardized: CRYSTALS-Dilithium (FIPS 204) and FALCON (FIPS 206) for signatures, ML-KEM (FIPS 203) for key encapsulation. The additional standardization track that HAWK was in is still in process, and this incident shows those candidates carry real risk.
For teams building systems with a long operational horizon, anything handling data with a 10-plus-year sensitivity window, the combination of harvest-now-decrypt-later attacks and AI-accelerated cryptanalysis compresses the margin for staying on unratified algorithms. Algorithm agility stops being a design nicety. Build your crypto layer so you can swap primitives without a full rewrite, because the review timeline between "candidate" and "broken" just got shorter.
The Research Pace Is the Real Story
What surprises me is not that AI found a flaw. Cryptographic weaknesses have always been found eventually. What surprises me is the compression. Two years of expert human review, then 60 hours for the model.
I have not run Mythos myself on cryptanalysis tasks. It is not in external preview as of this writing. But taking Anthropic's framing at face value, the model is doing something genuinely non-trivial in a domain that demands deep mathematical intuition. Cryptanalysis is not code synthesis. It is not text summarization. It requires building conceptual models of algebraic structures and looking for weaknesses in those models. The fact that a frontier model can now produce original published results in this space changes the shape of what AI is actually useful for in a research organization.
The practical question this raises: if AI can compress research timelines this sharply in cryptography, what other foundational assumptions in your stack are worth scheduling a second look at? I am not suggesting a panic review of everything. But if your threat model was "these primitives are safe because expert humans have reviewed them for years," that model just got revised.
Stay on the fully standardized NIST algorithms. Build for crypto agility. And pay attention to what Anthropic ships when Mythos becomes available, because this is a preview of what it does.
Comments
Post a Comment