Anthropic's Claude Mythos Preview identified a weakness in HAWK, a post-quantum encryption candidate, and improved attacks on reduced-round AES, in roughly 60 hours of work.
AI News
A post-quantum encryption candidate that survived two years of expert human review has a newly identified weakness, found by an AI model in 60 hours.
Anthropic disclosed on July 28 that its Claude Mythos Preview model had identified a flaw in HAWK, a digital signature scheme currently under evaluation by the US National Institute of Standards and Technology (NIST) as part of its post-quantum cryptography standardization process. The same model separately produced an improved attack on a reduced variant of the Advanced Encryption Standard (AES).
Neither finding affects any system currently running in production. HAWK has not been deployed, and the AES result applies only to a seven-round version of the cipher, not the full 10-round standard.
The HAWK Weakness
HAWK is one of several remaining candidates in NIST's effort to identify cryptographic systems that would remain secure against quantum computers. Quantum computers, if developed at sufficient scale, could break most of the encryption schemes currently in widespread use. HAWK had passed two full rounds of expert review over two years before Claude Mythos Preview found the flaw.
The weakness stems from a mathematical symmetry inside the lattice structure that underpins HAWK's security, a property known as a nontrivial automorphism. Prior research had established that finding such a symmetry would enable an attack but had not determined whether one existed in HAWK's specific lattice. Exploiting the symmetry cuts the scheme's effective key strength in half. Restoring the original security level would require doubling HAWK's key sizes, which undermines much of what made the scheme a competitive candidate in the first place.
One Anthropic researcher with a theoretical computer science background worked alongside the model during the discovery. That researcher had no specialist expertise in lattice-based cryptography. Claude Mythos Preview operated inside a multi-agent harness built on Claude Code, with access to Python, Sage, and published cryptographic literature. Human input was limited largely to project management decisions. The full process cost approximately $100,000 in API usage and took around 60 hours to complete.
Related Article: Here's How AI Can Help (and Hurt) Crypto Security
The AES Result
The AES finding followed a more automated process. A second researcher built a scaffold that allowed the model to work with minimal human direction. Claude Mythos Preview initially concluded that improving on existing AES cryptanalysis was not achievable and declined to engage with the problem. After a single prompt directing it to search for genuinely new approaches, it began producing results.
Over three days and approximately 1 billion output tokens, the model developed a technique it called the Möbius Bridge. The technique eliminates one of the guesses an attacker must make during a meet-in-the-middle attack on seven-round AES, producing an attack between 200 and 800 times faster than the previous best approach, depending on the measurement method applied. Two Anthropic researchers then spent several hundred hours validating the result, as neither had specialist cryptography expertise.
Anthropic shared its HAWK findings with the scheme's authors in June 2026 and coordinated disclosure to the public NIST mailing list. The company consulted with researchers at ETH Zurich, Tel Aviv University, and the University of Haifa throughout the process and briefed US government and industry partners before publication. Alongside the two primary findings, Anthropic released CryptanalysisBench, a benchmark designed to help track AI model performance on cryptanalytic tasks over time. The company said work on additional ciphers is ongoing and that further findings are expected to be published after validation is complete.
This article contains links to third-party websites or other content for information purposes only (“Third-Party Sites”). The Third-Party Sites are not under the control of CoinMarketCap, and CoinMarketCap is not responsible for the content of any Third-Party Site, including without limitation any link contained in a Third-Party Site, or any changes or updates to a Third-Party Site. CoinMarketCap is providing these links to you only as a convenience, and the inclusion of any link does not imply endorsement, approval or recommendation by CoinMarketCap of the site or any association with its operators. This article is intended to be used and must be used for informational purposes only. It is important to do your own research and analysis before making any material decisions related to any of the products or services described. This article is not intended as, and shall not be construed as, financial advice. The views and opinions expressed in this article are the author’s [company’s] own and do not necessarily reflect those of CoinMarketCap.
