Anthropic's Claude Mythos Preview Discovers New Cryptographic Attacks on HAWK and AES
Anthropic has released technical papers and chain-of-thought data detailing two new cryptographic attacks discovered by its Claude Mythos Preview model. These attacks target the HAWK post-quantum digital signature candidate and a reduced-round version of the Advanced Encryption Standard (AES). This research demonstrates the growing capability of AI models to autonomously discover complex cryptographic vulnerabilities, which could significantly impact future cybersecurity and AI safety red-teaming. While the findings do not compromise production systems, they highlight the need to evaluate AI capabilities in high-stakes security domains. The attack on HAWK significantly weakens the post-quantum digital signature scheme, while the attack on the reduced-round AES improves the speed of the previous best attacks by 200 to 800 times. Anthropic emphasized that these findings do not affect production systems or break the full, standard version of AES.
## BACKGROUND
HAWK is a digital signature scheme designed to remain secure against potential future quantum computer attacks, a field known as post-quantum cryptography. AES (Advanced Encryption Standard) is the global standard for encrypting sensitive data, and researchers often study "reduced-round" versions of it to understand its mathematical limits and robustness before attempting to break the full version.