Anthropic's AI Breaks HAWK-256 Cryptographic Challenge in 60 Hours

Anthropic researchers used an unreleased AI model to discover a practical key-recovery attack on HAWK-256, a cryptographic challenge parameter in post-quantum cryptography. The finding demonstrates that advanced language models can now identify vulnerabilities in mathematical primitives that have undergone extensive human review.

The HAWK Discovery

Claude Mythos Preview discovered the attack over approximately 60 hours, with one researcher providing project-management guidance rather than technical lattice cryptography expertise. The AI identified a previously unused symmetry—specifically, a nontrivial automorphism in the lattice structure—that reduces HAWK-256's effective security from 2^64 to 2^38 operations. The API cost totaled approximately $100,000.

HAWK-256 is a challenge parameter set for cryptanalysis, not a candidate for NIST standardization. The actual HAWK candidates under NIST review are HAWK-512 and HAWK-1024. The HAWK scheme is not deployed in any production systems, and the finding does not affect deployed infrastructure.

AES Breakthrough

The same research effort identified a new method to attack 7-round reduced AES-128, the most widely used symmetric cipher. The technique, which Anthropic calls the "Möbius Bridge," eliminates a 256-way guessing step and makes the attack 200 to 800 times faster than prior best-known methods. This discovery was almost fully autonomous after minimal human prompting over several days, generating approximately one billion output tokens.

The AES attack does not break full 10-round AES-128 and remains impractical, requiring roughly 2^105 chosen plaintexts. Nevertheless, it represents a significant theoretical advance in symmetric cryptanalysis.

Coordinated Disclosure

Anthropic coordinated disclosure with NIST, the HAWK authors, and government and industry partners before publication. The organization published implementation code on GitHub and partnered with ETH Zurich, Tel Aviv University, and the University of Haifa to release CryptanalysisBench, a 191-task benchmark for evaluating AI capabilities in cryptanalysis.

Why It Matters

The research signals a turning point in cybersecurity: AI systems can now autonomously discover vulnerabilities in complex mathematical constructions vetted by expert cryptographers. While neither finding compromises production systems, the work suggests that AI-assisted cryptanalysis could accelerate the discovery of flaws in emerging cryptographic standards.

As NIST continues standardizing post-quantum cryptography ahead of anticipated quantum computing threats, the Anthropic findings underscore the need for rigorous AI-assisted review alongside traditional human expert analysis. The organization's decision to publish both the methodology and benchmark suite indicates a commitment to transparency as the field grapples with AI's expanding role in both breaking and building cryptographic systems.