Anthropic's AI forces NIST to pull HAWK cipher after finding flaw humans missed for two years

Anthropic's Claude Mythos Preview has broken HAWK, a digital signature scheme under active consideration by NIST for post-quantum cryptographic standards, by discovering a structural flaw that human researchers spent two years failing to find. NIST withdrew HAWK from its post-quantum standardization process on July 29 following the disclosure.

The finding marks the first time an AI model has directly caused the removal of a cryptographic standard candidate — with significant implications for how the security community approaches both algorithm review and AI-assisted research.

The attack: a symmetry no one noticed

Claude Mythos found a nontrivial automorphism — a mathematical symmetry — buried in HAWK's lattice structure that preserves the lattice while dramatically reducing the cost of key recovery. For HAWK-256, the smallest configuration, the work factor for recovering a secret key dropped from 2⁶⁴ to 2³⁸ operations. That is roughly 67 million times less work. A 96-core server can now complete the attack in approximately three hours and 42 minutes.

The reduction scales to larger configurations: HAWK-512 drops from 2¹⁵⁰ to 2¹⁰⁸ equivalent operations, and HAWK-1024 from 2²⁸⁸ to 2¹⁸². The attack is exponential rather than polynomial — it does not break HAWK instantaneously — but the reduction is large enough that the scheme can no longer be considered secure at its stated parameters.

A second breakthrough: the Möbius Bridge

In a separate line of research, Claude Mythos invented a new cryptanalytic technique it named the Möbius Bridge, which accelerates an existing meet-in-the-middle attack against a weakened seven-round version of AES-128 by between 200 and 800 times depending on measurement methodology. The improvement works by identifying an invariant fingerprint that eliminates 256-way enumeration from the prior attack.

The Möbius Bridge result does not affect production AES systems. Seven-round AES is a reduced variant used in academic security analysis; full AES-128 has ten rounds and the attack requires approximately 2¹⁰⁵ chosen plaintexts, making it computationally infeasible in practice. Anthropic is clear that neither breakthrough endangers currently deployed encryption.

How the research happened

Anthropic's team spent roughly three days guiding Claude Mythos to the HAWK discovery, generating approximately one billion output tokens in the process. Human researchers then spent nearly one month verifying the result — verification, not discovery, was the primary bottleneck. Total API compute cost across both breakthroughs was approximately $100,000.

Notably, the model initially resisted the AES line of inquiry. Researchers reported that Claude Mythos "insisted that improving on AES was impossible" before persistent prompting produced the Möbius Bridge insight. Anthropic disclosed the HAWK vulnerability to the scheme's authors in June and coordinated public release simultaneously with notification to the NIST mailing list. HAWK's authors helped verify the result before disclosure.

What this means for post-quantum security

Governments and enterprises worldwide are racing to deploy post-quantum cryptography before quantum computers capable of breaking RSA and elliptic-curve schemes arrive. NIST's standardization process is the primary pathway for selecting algorithms that will protect sensitive data for decades. HAWK's withdrawal removes one candidate from that pool, but the other finalists — CRYSTALS-Dilithium, FALCON, and SPHINCS+ — were not affected and remain on track.

The more lasting impact may be on how cryptographic review is conducted. HAWK passed years of expert human analysis without the flaw being detected. As first reported by The Hacker News, Anthropic published the full implementation code, test vectors, and verification artifacts on GitHub, making the research reproducible for independent confirmation.