Bottom line: Claude Mythos Preview discovered a practically feasible attack on HAWK-256 and an accelerated method against reduced AES, both disclosed to NIST before publication, while the more standards-relevant HAWK-512/1024 remain practically unattackable.
Anthropic’s AI model Claude Mythos Preview has developed two new cryptanalytic attacks: one against the post-quantum signature scheme HAWK and one against a reduced AES variant. This is relevant for CISOs because HAWK is currently being standardised by NIST and potential weaknesses can be identified during the development phase.
Claude Mythos Preview has developed a previously unknown attack for complete key recovery against HAWK-256, a parameter set of the post-quantum signature scheme HAWK. HAWK is one of nine schemes that the US standards institute NIST included in the third round of its post-quantum signature selection process in May 2026. The attack exploits a previously unknown symmetry in the mathematical lattice and reduces the computational effort required for key recovery from approximately 2^64 to 2^38 operations. Anthropic’s implementation requires approximately three hours and 42 minutes on average for complete recovery on a server with 96 processor cores.
Anthropic emphasises that the attack affects only HAWK-256, not the relevant parameter sets HAWK-512 and HAWK-1024 in the NIST standardisation process. Anthropic also estimates reduced computational effort for these, but both variants remain practically unattackable.
The second finding concerns a version of AES-128 reduced to seven out of ten rounds. Such analyses on round-reduced variants are common in cryptanalysis to assess the security margin. Claude Mythos Preview developed a method called Möbius Bridge for this, which shortens an existing meet-in-the-middle attack by simplifying a guessing step. This accelerates the previously best known attack by 200 to 800 times. However, the attack requires approximately 2^105 chosen plaintexts under a fixed key and is therefore not practically applicable. The full ten-round AES-128 is unaffected by either finding.
For the HAWK attack, Claude Mythos Preview worked for approximately 60 hours in a multi-agent environment, supported by a researcher without cryptographic expertise. For the AES attack, the model required approximately three days and a total of approximately one billion output tokens. Anthropic put the API costs at approximately 100,000 US dollars each; the subsequent human verification was considerably more time-consuming – two researchers required nearly a month to verify the correctness of the AES method.
Anthropic disclosed both results to the HAWK developers as well as US government and industry partners before publication and coordinated the disclosure with NIST. It is not yet publicly known whether NIST or the HAWK developers plan changes to the scheme or its status in the standardisation process. At the time of publication, NIST continued to list HAWK as a third-round candidate.
Source: www.it-daily.net · Published 29 July 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.3.