An unreleased Anthropic model called Claude Mythos Preview has discovered a previously unknown vulnerability in HAWK, a lattice-based digital signature scheme designed to resist attacks from future quantum computers. The finding is notable not just for what it exposes in HAWK, but for who found it: human cryptographers had worked on the scheme for years without surfacing the flaw.
HAWK is currently one of the leading candidates in the U.S. National Institute of Standards and Technology's post-quantum signature competition. NIST advanced the scheme to the third round of evaluation in May 2026, where it stands as the last lattice-based candidate still in contention to become a federal cryptographic standard used across government and industry infrastructure.
A 67 Million-Fold Reduction in Attack Cost
According to the reporting, Claude Mythos identified a technique that slashes the computational cost of key recovery against HAWK's smallest configuration from roughly 2^64 operations down to about 2^38 — a reduction of approximately 67 million times less work than previously believed necessary. That gap is the difference between an attack considered computationally infeasible and one that becomes a realistic concern for a standard meant to protect data for decades.
NIST's own assessment of the scheme's third-round status reflects how far HAWK had advanced before this discovery surfaced: "NIST moved it into the third round of its post-quantum signature competition in May, where it is the last lattice-based candidate standing."
Related: Sandbox Escape Claude Cowork Menyusul Insiden OpenAI Seminggu Sebelumnya
What It Means for HAWK's Future
Fixing the vulnerability is possible, but the fix comes at a cost. Patching the flaw would require doubling HAWK's key sizes, which undercuts the scheme's core selling point: compact signatures that keep block space and transaction fees low in systems that rely on them, including blockchain protocols. That tradeoff could weaken HAWK's competitive position against rival post-quantum candidates still under NIST review.
AI as Both Discovery Tool and Bottleneck Risk
The episode adds to a growing body of evidence that frontier AI models can now outperform specialist researchers at certain classes of cryptanalysis. It also raises a less comfortable implication: as models like Claude Mythos get better at finding these flaws, verifying and responding to AI-discovered vulnerabilities may increasingly strain the limited pool of humans qualified to check the work. As one researcher put it, human researchers may become the bottleneck in addressing what AI systems are capable of finding.
No deployed cryptocurrency or financial system is affected by the HAWK finding directly, since the scheme has never been used commercially. But the discovery lands at a moment when Anthropic's own products have already drawn security scrutiny — the company recently had to respond to a sandbox escape incident in its Claude Cowork product, underscoring that AI labs are now navigating security stories on two fronts: vulnerabilities their models find, and vulnerabilities found in their models.
