Claude Helps Find Cryptographic Weaknesses, Signaling a New Role for AI in Cryptanalysis

Introduction

Anthropic has reported that Claude Mythos Preview helped researchers uncover weaknesses at the level of cryptographic algorithms themselves. The two headline findings involve HAWK, a post-quantum digital signature candidate, and a reduced-round version of AES, the widely used symmetric cipher. Anthropic stresses that neither result has a practical impact on deployed systems today, but both show that frontier AI models may become meaningful tools in cryptanalysis.

Key points

• A stronger attack on HAWK: HAWK is a third-round candidate in NIST’s additional call for post-quantum digital signature schemes. According to Anthropic, Claude Mythos Preview helped improve the best-known attack in about 60 hours of work, effectively cutting the scheme’s key strength in half. The attack relies on identifying a previously unexploited nontrivial automorphism in the lattice structure used by HAWK, enabling faster enumeration.
• A faster attack on reduced-round AES: AES itself remains unbroken in this work. Researchers often study weakened versions of AES to understand the cipher’s security margin. Mythos found a way to attack one such reduced-round version by eliminating one of the guesses an attacker would otherwise need to make, improving on previous attacks by roughly 200 to 800 times.
• Mostly autonomous, but still guided research: The HAWK result came from collaboration between one Anthropic researcher and Claude, while the AES finding used a scaffold that allowed Claude to search more autonomously. Anthropic says each result cost around $100,000 in API usage to develop.
• Responsible disclosure and benchmarking: Anthropic says it consulted academics, shared the HAWK attack with its authors, coordinated disclosure to the public NIST mailing list, and briefed government and industry partners. It also partnered with ETH Zurich, Tel Aviv University, and the University of Haifa to build CryptanalysisBench for evaluating LLMs on cryptanalytic tasks.

Why it matters

The important point is not that current encryption has suddenly failed. HAWK is still a candidate and is not deployed, while the AES result targets a deliberately weakened version rather than the full standard. Instead, the significance lies in the shift from finding coding mistakes to assisting with mathematical discovery.

For cryptography, this is both a warning and a useful stress test. More capable AI systems could lower the cost of finding weaknesses in algorithms, including candidates under review. At the same time, using AI to pressure-test designs before deployment could make future standards stronger.

For the AI field, the research illustrates a broader governance challenge. Models that can advance scientific work may also expand offensive capabilities in security-sensitive domains. The next phase will require careful benchmarking, responsible disclosure norms, and clearer boundaries around how such capabilities are developed and shared.

Source: Hacker News