AI-Enhanced Cryptanalysis: Claude Mythos Reveals Full Attack on HAWK-256 and Accelerates 7-Round AES-128

AI-assisted cryptanalysis by Claude Mythos reveals a full attack on HAWK-256 and accelerates 7-round AES-128. Learn about its impact on cybersecurity.

AI-Enhanced Cryptanalysis: Claude Mythos Reveals Full Attack on HAWK-256 and Accelerates 7-Round AES-128
Vulnerabilities

Illustrative image generated with AI

Introduction

On July 28, 2026, a group of researchers released two cryptanalytic results obtained with the assistance of Anthropic's Claude Mythos Preview model. In a multi-agent environment, over approximately 60 hours and with an estimated API cost of $100,000, the AI helped derive a key-recovery attack against the HAWK-256 digital signature scheme – a challenge parameter – and a 200‑ to 800‑fold improvement for a known attack on 7‑round reduced AES-128. Code, data, and papers were publicly released to ensure reproducibility. The results do not pose an immediate danger to real‑world systems, but they raise questions about the robustness of security estimates and the role of artificial intelligence in cryptographic research.

Technical Analysis

HAWK and the Lattice Automorphism

HAWK is a lattice-based signature scheme that entered the third round of the NIST post‑quantum standardization process in May 2026. The attack exploits a previously unused lattice automorphism, reducing the smLIP (shortest modular Lattice Isomorphism Problem) to a search for short vectors in a half‑dimensional lattice. The provided implementation recovers equivalent signing material in approximately 3 hours and 42 minutes on 96 cores, lowering the work factor for HAWK‑256 from 2⁶⁴ to 2³⁸.
For the standard HAWK‑512 and HAWK‑1024 parameters, gate count estimates drop to 2¹⁰⁸ and 2¹⁸², respectively, remaining impractical. The attack does not constitute a polynomial break nor does it extend to other post‑quantum schemes.

7‑Round AES‑128 and the “Möbius Bridge”

On the symmetric encryption front, the model constructed an invariant footprint called the “Möbius Bridge,” which eliminates a 256‑way enumeration step from a previous meet‑in‑the‑middle attack. The acceleration applies only to the 7‑round variant and remains purely theoretical: it requires about 2¹⁰⁵ chosen plaintexts. Full 10‑round AES‑128 is not affected in the slightest.

The methodology demonstrates how generative AI can identify novel mathematical properties and optimize existing attack vectors, opening new directions for assisted cryptanalysis.

Impact

No immediate modifications are required for production systems or software. The HAWK attack is limited to the challenge parameter HAWK‑256, while the versions intended for practical use maintain wide safety margins, albeit with downward‑revised estimates. For AES, the data and computational complexity remain prohibitive in any real‑world scenario.
The most concrete impact is indirect: HAWK security evaluations may be updated within the NIST process, potentially leading to re‑calibration of parameters or resilience claims. From a scientific perspective, the results show how AI can accelerate the discovery of vulnerabilities, suggesting a paradigm shift in cryptanalytic research.

Mitigation

  • HAWK: standard parameters (512, 1024) remain out of reach of currently known attacks. No urgent action is needed; any adjustments will follow the official NIST process.
  • AES: no countermeasures are required, as the attack concerns only a laboratory variant.
  • Recommendations: monitor developments in the NIST process and communications from research teams, avoiding hasty measures. Existing implementations can continue to operate according to current best practices.

FAQ

  1. Does the HAWK‑256 attack compromise all HAWK schemes?
    No. The full attack is limited to the challenge parameter HAWK‑256, designed specifically to test robustness. Standard HAWK‑512 and HAWK‑1024 parameters, although showing reduced security estimates, remain computationally unassailable with current technology.

  2. Is AES‑128 now considered insecure?
    Absolutely not. The optimization concerns only the reduced 7‑round version, already known and with no practical relevance. Standard 10‑round AES, used in countless applications, is not threatened by this result.

  3. Can artificial intelligence break cryptography autonomously?
    No. Claude Mythos assisted researchers in identifying mathematical optimizations, but the process was guided by human experts in a multi‑agent environment. AI does not have the ability to conduct cryptographic attacks on its own, but it proves to be a powerful tool for accelerating analysis and inspiring new research directions.

Security dossiers

Read next

Sources

This article is an original reworking based on the sources below.

Back to home

Latest Cybersecurity News

All cybersecurity news →