Claude Mythos cracks post-quantum crypto math

For years, breaking encryption usually meant catching human programmers slipping up on an implementation bug. Now, AI is going after the pure math itself.

Anthropic just revealed that its Claude Mythos Preview model uncovered major mathematical weaknesses in two core cryptographic algorithms. The model halved the security of HAWK, a post-quantum signature candidate currently under evaluation by NIST, and massively sped up attacks against a round-reduced version of AES, the gold standard of symmetric encryption.

A $100K API bill versus two years of human review

HAWK was engineered to keep communications safe from future quantum supercomputers. It had already survived two intense rounds of review by elite human cryptographers over two years, making it into the third round of NIST's standardization race.

Anthropic paired Claude Mythos with a single researcher who had a general computer science background but no deep expertise in lattice cryptography. Claude operated in a multi-agent harness equipped with Python, Sage, and access to math papers. Interestingly, when one agent prematurely discarded a promising mathematical lead as a dead end, a second agent picked it up, kept running, and proved the exploit.

Within 60 hours and roughly $100,000 in API credits, Mythos spotted an unexploited nontrivial automorphism in HAWK's lattice structure. That breakthrough sliced its effective key size right down the middle: cracking HAWK-256 used to take 2^64 operations, but Claude dropped it to 2^38. While it is still an exponential attack, fixing it means doubling HAWK's key size, which pretty much destroys the lightweight appeal that made it a top NIST contender in the first place.

On the AES front, Anthropic let Claude run completely autonomously inside a testing harness. The model eliminated an entire guessing step previously required by cryptanalysts, improving attack speeds on round-reduced AES by 200x to 800x for another $100,000 in compute. To help researchers evaluate similar capabilities, Anthropic partnered with ETH Zurich, Tel Aviv University, and TU Berlin to launch CryptanalysisBench.

Why it matters

To be clear, you do not need to scramble to change your passwords today. HAWK is merely an experimental candidate that was never deployed to real users, and full-round AES remains rock solid. Anthropic handled disclosure by the book, notifying HAWK's authors back in June and sharing details with NIST.

What matters here is the shift in asymmetric warfare. Designing cryptographic standards has traditionally been an excruciating, decade-long slog where human mathematicians try to break each other's work by hand. Claude Mythos just proved that an AI agent team, steered by someone who isn't even a domain specialist, can outmaneuver years of peer review in a single weekend.

We think this fundamentally changes how security standards will be built. In the near future, submitting an encryption scheme without first subjecting it to hundreds of hours of autonomous AI red-teaming will look downright irresponsible. The bright side is that AI can help catch subtle mathematical flaws before algorithms hit production. The catch, of course, is that a $100,000 API bill is well within reach for nation-states and well-funded threat actors looking for zero-days in the math we all depend on.