Claude Mythos cracks post-quantum candidate HAWK-256, speeds AES-128 attack

In a result that made waves on Hacker News (114 points) and in security circles, Anthropic reported that its Claude Mythos Preview model contributed to an end-to-end key-recovery attack against HAWK-256, exploiting a newly identified symmetry in the underlying lattice. The estimated runtime is about three hours and 42 minutes on a 96-core server — practical enough that the HAWK team withdrew their scheme from NIST's post-quantum standardization process after confirming the attack's impact.

Separately, the model accelerated an attack on seven-round AES-128 by a factor of 200 to 800. Anthropic was careful to note that neither result affects production systems: full AES uses more rounds, and HAWK was a candidate, not a deployed standard. Still, the demonstration is significant because it shows a frontier model contributing novel cryptanalysis rather than merely reproducing known techniques.

The mechanism — an AI surfacing a lattice symmetry that human cryptographers had not exploited — is the story's core. It reframes AI cryptanalysis from theoretical possibility to concrete tool, and it is the kind of dual-use capability that fed directly into the same week's debate over whether labs should pace development.

Context and caveats: the HAWK withdrawal is a real, verifiable outcome (a scheme leaving NIST's process), which lends credibility versus a lab-reported benchmark. Skeptics will want independent replication of the AES speedups, and cryptographers note that candidate schemes being broken during standardization is precisely what the NIST process is designed to catch — arguably a success of the process, not a crisis. What to watch is whether similar techniques generalize to deployed standards.