Anthropic AI Discovers Mathematical Flaws in Cryptographic Algorithms

Anthropic announced on July 28, 2026, that its Claude Mythos Preview AI model discovered fundamental mathematical weaknesses in two cryptographic algorithms. The model identified a structural flaw in HAWK, a post-quantum digital signature scheme currently under review by the National Institute of Standards and Technology. Within 60 hours of semi-autonomous work, the AI identified a nontrivial automorphism that effectively halved HAWK's key strength, reducing attack costs for the HAWK-256 variant from 2^64 to 2^38 operations.

Additionally, the model developed a new attack technique called the Möbius Bridge targeting a seven-round reduced version of the Advanced Encryption Standard (AES). This technique improved the speed of previous best attacks on the AES-128 cipher by 200 to 1,000 times. Each discovery cost approximately $100,000 in API compute. While neither finding compromises production systems—as HAWK is not yet deployed and the AES attack targets a reduced version—the results demonstrate that AI can now perform cryptanalysis at a depth that outpaces years of human expert review.

Anthropic coordinated the disclosures with NIST and HAWK's authors in June 2026 and subsequently launched CryptanalysisBench, a partnership with ETH Zurich, Tel Aviv University, and the University of Haifa to evaluate AI capabilities. These developments have intensified concerns regarding the security risks of agentic AI.