Security Current Affairs

Claude AI Cracked a Post-Quantum Standard in 60 Hours. Humans Couldn't for Years.

2026-07-29

On July 28, 2026, Anthropic announced that Claude Mythos Preview—an unreleased model—had done something human cryptographers failed to do: found a working attack against HAWK-256, a digital signature scheme that survived multiple rounds of expert review inside NIST's post-quantum standardization process.

The vulnerability is not trivial. Mythos spent roughly 60 hours semi-autonomously (at about $100,000 in API costs) and effectively cut HAWK's key strength in half. The attack exploits a previously unused symmetry in the scheme's lattice structure, reducing the operation count from ~2^64 to ~2^38. That's a massive gap between "theoretically possible someday" and "actually broken."

Why does this matter for crypto? HAWK was chosen as one of nine candidates in NIST's third round of post-quantum digital-signature evaluation partly because of its compact signatures—shorter than Falcon, lower blockchain overhead, fewer transaction fees. Doubling key sizes to restore security erases that entire selling point. For Bitcoin and Ethereum long-term security against quantum threats, this matters.

But Mythos didn't stop there. The model also invented a new attack technique it named the Möbius Bridge, making an existing assault on a 7-round variant of AES between 200 and 800 times faster. Again: research variant only. The real AES-128 uses ten rounds and remains secure. Still, the pattern is unsettling. An AI systematically broke things humans had certified as safe.

Then came LEA, a Korean national and ISO standard for lightweight encryption on mobile devices. Mythos developed a practical attack recovering 13-round LEA keys in under 2^30 encrypted plaintexts, running in under an hour on a modern desktop. The full 24-round cipher is untouched, so this has no immediate threat. But the tally mounts.

What's actually interesting here is the labour split. Anthropic gave Mythos direction, computing resources, and verification—then spent several hundred hours learning enough cryptography to confirm the results actually worked. Two technical papers and reproducibility artifacts went public. That's not some mysterious black-box result. That's auditable science.

Anthropic followed responsible disclosure protocols. NIST was notified. Algorithm authors, government and industry partners all got advance warning before publication. No surprise drops. No scorched-earth announcements.

Yet there's a deeper tension the whole incident exposes. The bottleneck in cryptographic standardization is no longer discovery—it's verification. As large language models become capable of systematic cryptanalysis, human experts now face a triage and remediation problem they're not equipped to handle at scale. The vulnerability-finding engine is running faster than the human processes that validate and fix things.

For post-quantum cryptography, already years behind the quantum threat timeline, this could accelerate NIST's standardization process. Or complicate it. Probably both. When an AI can crack something in two days that humans missed for years, the entire evaluation framework gets questioned. HAWK survives if key sizes double—but was doubling key sizes part of the original plan? And what else hasn't been tested hard enough yet?

The cybersecurity community is grappling with this now. Bug discovery is no longer the constraint.

Source & further reading:

Sources