Claude Mythos Discovers Flaw in HAWK Post-Quantum Signature Algorithm
Anthropic's Claude Mythos Preview model semi-autonomously discovered a critical vulnerability in the HAWK post-quantum signature algorithm, leading to its immediate withdrawal from the NIST standardization process. The AI model also successfully optimized attacks on a simplified version of the AES-128 encryption standard. This breakthrough demonstrates the growing capability of large language models in advanced cryptanalysis, showing they can find flaws in complex cryptographic candidates that human experts missed. It highlights how AI could accelerate both the vulnerability discovery and the validation processes for future security standards. The attack on HAWK exploited a previously undiscovered symmetry in its underlying lattice structure, halving its effective key strength and reducing the computational cost for HAWK-256 key recovery from 2⁶⁴ to 2³⁸. The discovery cost approximately $100,000 in API usage, and the vulnerability was confirmed by NIST within an hour of disclosure.
## BACKGROUND
Post-quantum cryptography (PQC) focuses on developing cryptographic algorithms that are secure against both quantum and classical computers, as quantum computers could easily break current standards like RSA or ECC. NIST has been running a multi-year standardization process to select and evaluate these new algorithms, many of which rely on hard mathematical problems in lattice-based cryptography.