Anthropic AI uncovers new attacks on HAWK and AES

Anthropic reported in July 2026 that an unreleased version of its Claude Mythos Preview model found previously unknown attacks on several cryptographic algorithms, including HAWK and a seven‑round research version of AES‑128.

The model identified a mathematical symmetry in HAWK, a lattice‑based digital‑signature candidate that NIST moved to the third round of its post‑quantum signature competition. For HAWK’s smallest configuration the work to recover a secret key fell from 2^64 operations to 2^38. Anthropic noted that fixing the weakness would require roughly doubling HAWK’s key sizes and that HAWK has never been deployed in the field.

For AES‑128 the team constrained the model from using known classes of cryptanalysis and asked it to invent a new approach. Claude produced what the paper calls a “Möbius Bridge,” an idea that removes one of nine key bytes an attacker had to guess in the seven‑round research variant. After refinement, the new attack ran between 200 and 800 times faster than the prior best result for that seven‑round version.

Anthropic also reported a faster attack on LEA, a Korean national and ISO lightweight cipher. The model produced a method to break 13 rounds in under an hour on a desktop, compared with a prior published result that required 2^98 plaintext pairs. Deployed LEA uses 24 rounds, so the finding does not break the cipher as used in practice.

The company coordinated disclosure of the HAWK result with NIST and notified algorithm authors, U.S. government partners and industry contacts before publishing details. Anthropic published transcripts and technical material showing the model’s interaction with researchers. Each of the published results incurred roughly $100,000 in API usage, and researchers spent several hundred hours confirming the AES attack.

Anthropic built CryptanalysisBench, a set of 191 cipher‑breaking tasks drawn largely from NIST competitions, to benchmark models. In those tests a later Mythos model solved 85.7% of tasks that already had known solutions, compared with 65.3% for the weakest model tested. For full‑strength ciphers with no published breaks, every model scored under 9%.

The paper states: “The majority of mathematical discoveries in this paper were AI‑assisted. Human author contribution mainly consisted of directing, organizing and verifying AI work.” Transcripts show the model initially resisted simpler approaches and continued iterating until it produced new attacks.

Anthropic warned that the speed and volume of AI‑discovered vulnerabilities could exceed traditional human processes for triage, verification and remediation, and that human researchers may become the bottleneck in responding to automated cryptanalytic discoveries.