The model Mythos from Anthropic found weaknesses in the cryptographic schemes HAWK and AES

Anthropic said that its model Claude Mythos Preview discovered mathematical weaknesses in two cryptographic schemes: it improved an attack on the post-quantum signature scheme HAWK and developed a new attack on a weakened version of the AES-128 cipher (a version with 7 of 10 rounds). According to Anthropic, neither finding threatens systems deployed in production – HAWK is only a candidate in the third round of the NIST standardization process, and the tested version of AES is not the one commonly used.

For HAWK, the model worked largely independently in a multi-agent system and found an improved attack based on a previously undetected symmetry in the mathematical lattice in 60 hours, while human experts had been studying the scheme for over two years. One agent initially deemed the idea infeasible, but another fully exploited it. According to Anthropic, a human researcher with a theoretical background in computer science, but no expertise in lattice-based cryptography, focused mainly on project management. For AES, the model initially refused the task, saying that further improvement was impossible, and only after being prompted to look for “truly new ideas” did it develop a method called Möbius Bridge, which, according to Anthropic, improves on the best existing attacks by a factor of 200 to 800. Over three days, the model generated roughly a billion tokens and received only three substantive instructions from a human. API costs for the two tasks reached roughly 100 000 dollars; human researchers without cryptographic expertise then spent hundreds of hours verifying the results.

Ars Technica adds important caveats to the results: these are incremental improvements, not a breach of systems in use; the tested variants are deliberately weakened “challenge” versions intended for independent assessment, and the methods used would probably not be feasible outside the testing environment. According to Ars Technica, it is not yet possible to distinguish precisely how much of the presentation from Anthropic is marketing and how much is a substantive result.

Anthropic shared the results in advance with the US government, industry partners and the authors of the HAWK scheme. Together with researchers from ETH Zurich, Tel Aviv University and the University of Haifa, it also created the CryptanalysisBench benchmark to systematically evaluate the cryptanalytic capabilities of language models. The model Mythos Preview remains unavailable to the public; according to Ars Technica, only a limited group of trusted users has access to it.