IA in the Mythos Preview cryptoresearch adjusts the safety margins of HAWK-256 and AES-128 without affecting production

Anthropic has announced two cryptanalogical results obtained with the help of its Mythos Preview research model that, although they do not break production systems, modify the assessment of the safety margin in specific scenarios: an end-to-end key recovery against the HAWK-256 challenge parameter and a substantial acceleration of an attack against AES-128 reduced to seven rounds. Both findings are technically relevant to the cryptographic community, but none require immediate changes in production software.

In the case of HAWK, the key contribution was the identification of a additional automation in the network structure (lattice) that allows to reduce the cryptanalitic search to a problem with a significantly smaller dimension. This reduction, articulated by Anthropic as a construction of the so-called τ-cocycle lattice followed by network reduction and sifting, resulted in a public implementation that, according to the company itself, recovers an equivalent functional key for HAWK-256 in an estimate of about three hours and 42 minutes on a 96-core server. It is important to stress that HAWK-256 is a challenge parameter used as a target for analysis, not the configuration NIST considers for its safety levels (HAWK-512 and HAWK-1024); therefore, the versions that seek to offer practical resistance remain, for now, out of realistic reach of the attack.

Anthropic quantifies the impact: a reduction in expected work for HAWK-256 from 2 ^ 64 to 2 ^ 38 and adjustments in quantum cost estimates (gate-count) for higher parameters, although both major sets remain impractical to attack. The technique does not constitute a failure in polynomic time or a general weakening of network-based cryptography, and the authors warn that the improvement does not automatically extend to other candidates or cryptographic families.

The work on AES-128 is of a different nature: it studies residual safety against attacks on reduced versions of the algorithm - a standard practice for measuring margin - and presents the so-called Möbius Bridge construction, an invariant that allows to remove a listing of 256 possibilities in an existing meet-in- the-middle attack. Although this produces an estimated acceleration of 200 to 800 times at the affected stage, the complete vector for seven rounds continues to require a number of completely impractical chosen texts (of the order of 2 ^ 105), so no operational risk over AES in real 10-round implementations.

Beyond the technical details, the institutional part of the story is relevant: Anthropic published technical articles and reproducible artifacts and states that Mythos Preview generated most of the discoveries, with human orientation, computation resources and an extensive verification phase. The company has also published code that attacks exclusively HAWK-256 and contains public test keys; however, at the close of this review there is no confirmed independent reproduction of the complete attack on HAWK-256 in the public register. The absence of independent reproduction is a critical point that the community must address.

For the NIST post-quantum standardization process, in which HAWK was the only network-based scheme led to the third additional round in May 2026, these results modify the risk assessment on challenge parameters and stress the importance of continuous analysis. NIST and HAWK proposers are likely to review and discuss these figures; however, at the moment there is no public indication of changes in the status of the scheme.. The NIST project documentation can be followed here: https: / / csrc.nist.gov / projects / post-quantum-cryptography.

For system managers, product managers and security equipment, the practical recommendation is simple: there is no need for emergency measures. The production implementations do not depend on HAWK-256, and the variants with NIST-led parameters maintain their operational recommendation. However, it is appropriate to monitor the statements of the HAWK and NIST team, to review the reproducibility of the published code and, in post-quantum transition scenarios, to prefer parameters with established and well-audited margins.

For researchers and cryptographic risk assessment teams, the case is instructive on two fronts: first, it exemplifies how new algebraic symmetries can open attack routes that previously seemed closed; second, it shows the increasing relevance of advanced language models as assistants in complex research. This implies a double obligation: to exploit the creative opportunities offered by these models and, simultaneously, to strengthen verification practices, peer review and experimental independence. Anthropic has disseminated prompts, iterations and cost measures (along with an estimated cost of API close to six digits), which facilitates external audit; the community should prioritize reproduction and scrutiny.

In terms of policy and governance, the episode suggests that standardization agencies should incorporate explicit mechanisms to assess IA-assisted results, demand reproducible devices and promote transparency in verification processes. Cryptography remains based on evidence and reproducibility: discoveries assisted by IA do not change that principle, only accelerate the need for rigorous controls.

Since Anthropic has made its materials and code available to the public, those who wish to examine the artifacts and publications can start with the official pages of the company and its repositories: https: / / www.anthropic.com / research and https: / / github.com / anthropic. The academic and engineering community should replicate, extend and test the claims; independent confirmation is the step that will transform an interesting finding into a change of policy or parameters.

In summary, the work reported by Anthropic represents useful advances in the assessment of safety margins and shows the capacity of large models as advanced research tools. However, the conclusions are limited: they affect a challenge parameter and a reduced version of AES, do not derail the recommended schemes or require immediate patches in production systems. The practical lesson is clear: monitoring, verification and dialogue between standardization teams, authors of proposals and the research community to translate technical findings into informed decisions.