Security Current Affairs

Claude Mythos Broke Post-Quantum Crypto in 60 Hours. Humans Had Two Years and Failed.

2026-07-29

Anthropic's Claude Mythos Preview just did something properly unhinged: it found two previously unknown attacks on cryptographic algorithms, one of them targeting a scheme literally being considered as a U.S. federal standard for quantum-resistant signatures.

The AI cracked HAWK in 60 hours. Two years of global expert review, gone. HAWK is a digital signature system built to survive quantum computers—the whole point being that it works now, before the quantum threat materialises. It was supposed to be safe. It isn't.

Here's the mechanics. Mythos found a specific, previously unexploited symmetry in the lattice that HAWK uses. A nontrivial automorphism, if you want the technical term. For the smallest HAWK variant, the expected cost of a full key recovery attack dropped from 2⁶⁴ operations to 2³⁸. That's roughly 67 million times faster. On a 96-core server, Anthropic's code recovers HAWK-256 signing material in three hours and 42 minutes flat.

Why does this matter beyond the obvious "oh no our standards are broken" angle?

Signature size is block space. Block space is fees. Any blockchain hunting for a quantum-resistant replacement isn't just picking security—it's partly shopping by bytes per signature. HAWK's entire pitch was compact keys and fast signing. The fix costs it that edge entirely. It just became a lot less attractive.

The second finding targets a reduced-round variant of AES, the world's most widely used encryption cipher. Claude Mythos Preview developed a new fingerprinting technique called "Möbius Bridge" that improved an existing meet-in-the-middle attack, delivering a 200- to 800-fold speedup against seven-round AES-128. But here's the important bit: it only applies to seven of AES-128's ten rounds, still requires an impractical number of chosen plaintexts, and remains exponential. It doesn't extend to other NIST candidates or lattice cryptography generally. So full AES isn't broken. But the proof of concept is brutal.

The whole thing happened over roughly 60 hours in a multi-agent environment. Anthropic threw one human researcher at it for occasional project management, not even a lattice-cryptography specialist. The API cost about $100,000. Anthropic staff spent several hundred hours afterward just verifying the work was real.

What's wild is the asymmetry. Checking the work is now the slow part. Humans cannot keep pace with the model.

HAWK has never been deployed anywhere, and Bitcoin still runs on ECDSA, the pre-quantum scheme that candidates like HAWK are supposed to eventually replace. So the immediate risk is contained. Anthropic disclosed the HAWK attack to its authors in June and coordinated public disclosure to the NIST mailing list. Professional. Responsible.

But the meta-point here is unavoidable: an AI system just did research-level cryptanalysis that had defeated exhaustive peer review from actual experts over two years. It found a mathematical flaw nobody else saw. It did this faster, cheaper, and better than the humans who'd been staring at the problem since the thing was submitted.

That's not a small thing for the cryptography community. Or for anyone betting their long-term security on standards they assumed were solid.

Source & further reading:

Sources