Claude AI Just Cracked a Post-Quantum Test Scheme and Found a Faster 7-Round AES Attack

Anthropic has announced significant advancements in cryptographic research using their Claude Mythos Preview, achieving breakthroughs in attacks on HAWK-256 and AES-128 encryption schemes. The company developed an end-to-end key-recovery attack against HAWK-256 and significantly accelerated an attack on a reduced version of AES-128. Despite these achievements, Anthropic emphasized that these findings do not impact current production systems, as HAWK remains a candidate in the NIST post-quantum standardization process, and the AES attack requires impractical conditions for real-world application. The attack on HAWK-256 exploits a previously unused symmetry in the lattice structure of the signature scheme. Anthropic's implementation of this attack can recover keys in approximately three hours and 42 minutes on a 96-core server. The attack leverages a nontrivial automorphism that reduces the problem to finding a short vector in a lattice of half the original dimension, a path initially opened by researchers Daniël van Gent and Ludo Pulles. Anthropic's approach constructs a τ-cocycle lattice from the public key, using lattice reduction and sieving to recover short vectors and reconstruct a secret basis for signing messages. Anthropic's released code for the HAWK-256 attack supports only this specific parameter and rejects other inputs. The repository includes public keys that have been successfully attacked and supports the generation and testing of fresh HAWK-256 keys. The estimated work factor for HAWK-256 key recovery has been reduced significantly, but larger parameter sets like HAWK-512 and HAWK-1024 remain impractical to attack. The attack remains exponential and does not constitute a polynomial-time break, nor does it extend to other NIST signature candidates or lattice cryptography in general. The AES-128 attack focuses on a reduced version of the encryption scheme, targeting seven of its ten rounds. This attack assumes an adversary can obtain an impractical number of chosen plaintexts, making it irrelevant for real-world scenarios. The attack improves upon previous meet-in-the-middle methods by introducing an invariant fingerprint called the Möbius Bridge, which removes the need for a 256-way guessing step, resulting in a 200- to 800-fold speedup. Anthropic's research involved extensive verification and development over approximately 60 hours in a multi-agent environment, with human researchers providing guidance and verification. The cost of the API usage for this research was around $100,000, with significant human effort required for verification. The AES-128 attack remains theoretical, with Anthropic projecting the complete attack's cost from component measurements rather than executing a full recovery. The research findings were published alongside technical papers and reproducibility artifacts, contributing to the broader cryptographic research community. Anthropic's disclosures follow the release of CryptanalysisBench, a benchmark developed in collaboration with several universities to evaluate cryptographic models. The benchmark assessed Mythos 5, an update to Mythos Preview, which was instrumental in the HAWK and AES research. As of the latest updates, HAWK remains a candidate in the NIST standardization process, with Anthropic acknowledging the HAWK team's assistance in verifying results. The public thread on the NIST forum contained no replies, and independent reproduction of Anthropic's HAWK-256 recovery had not been reported at the time of review. The research underscores the potential of advanced AI models in cryptographic analysis, while also highlighting the challenges of translating theoretical breakthroughs into practical applications.