Researchers at Anthropic have used their Claude Mythos Preview model to discover cryptographic weaknesses in two important algorithms: HAWK, a post-quantum digital signature scheme under NIST review, and a reduced-round variant of the Advanced Encryption Standard (AES). These are research findings with no immediate impact on production systems, but they demonstrate frontier AI models’ ability to find mathematical flaws that have survived years of expert human scrutiny.

For HAWK, Claude Mythos Preview worked semi-autonomously over 60 hours with occasional human guidance to develop an attack that reduces the scheme’s effective keysize by a factor of two. The attack exploits a previously unexploited symmetry (a nontrivial automorphism) in the lattice underlying HAWK, enabling a faster exponential-time enumeration attack. While still exponential, the improvement means HAWK keys must be doubled to maintain the same security level, undermining many of the scheme’s advantages as a post-quantum candidate. The discovery cost roughly $100,000 in API credits. The attack does not apply to other NIST post-quantum candidates or lattice-based cryptography in general.

For AES, Claude Mythos Preview autonomously discovered a new meet-in-the-middle attack on a 7-round version of AES-128 (the full cipher has 10 rounds). The key innovation is a novel fingerprinting technique the model called a “Möbius Bridge,” which eliminates one of the attacker’s guesses (256 values) by using a transform invariant to that guess. Combined with other optimizations, the attack is 200 to 800 times faster than previous best attacks on the same reduced-round variant. The model worked nearly autonomously for three days, generating about one billion output tokens, with only three substantive human prompts. Human researchers then spent weeks validating the results. The attack operates under a chosen-plaintext threat model and does not break the full 10-round AES.

Beyond these two main results, Claude Mythos Preview has also found practical full key-recovery attacks on 13 rounds of the Lightweight Encryption Algorithm (LEA) and on 6 rounds of the Serpent-128 cipher, as well as limited improvements on attacks against Salsa20, Poseidon, and SHA-1. These are preliminary and require further validation. Anthropic has partnered with ETH Zurich, Tel Aviv University, and the University of Haifa to build CryptanalysisBench, a benchmark for evaluating LLM cryptanalytic capabilities.

The article emphasizes that these findings are the strongest attacks discovered to date, but they follow standard responsible disclosure practices, including coordination with the HAWK authors and NIST. The authors argue that as language models increasingly produce novel cryptographic research, human researchers may become bottlenecked on verification rather than discovery. They see this as an opportunity to stress-test a broader range of ciphers, including those that have received less scrutiny. The post concludes by noting the need for community discussion on how to handle future discoveries that might have immediate real-world impact.