Claude Mythos Preview Cryptanalysis Research

AI-Driven Cryptanalysis: Discovering Algorithmic Weaknesses

Claude Mythos Preview has demonstrated the ability to identify mathematical flaws within cryptographic algorithms themselves, moving beyond the discovery of implementation errors in code. In two primary research efforts, the model improved known attacks on the HAWK post-quantum digital signature scheme and a reduced-round version of the Advanced Encryption Standard (AES). While these findings do not currently impact production systems, they signal a shift in the ability of frontier AI models to perform expert-level cryptanalysis.

Improved Key Recovery Attack on HAWK

Claude Mythos Preview discovered an attack that effectively halves the key strength of HAWK, a digital signature scheme currently being considered by the National Institute of Standards and Technology (NIST) for post-quantum cryptography (PQC) standardization.

Technical Mechanism

HAWK's security relies on the Lattice Isomorphism Problem. Mythos identified a previously unexploited symmetry—a nontrivial automorphism—within the lattice used by HAWK. While prior academic work had established that finding such an automorphism would enable an attack, it remained unknown if one was accessible in HAWK's specific lattice. Mythos's discovery enables a faster enumeration attack; although the attack remains exponential in time, it necessitates doubling the key size to maintain the original security level.

Discovery Process and Cost

An Anthropic researcher worked semi-autonomously with Mythos in an agentic harness equipped with Python and Sage. The process involved an extensive literature review, mathematical reasoning, and computational experiments. The discovery and verification of the attack took approximately 60 hours and cost an estimated $100,000 in API credits.

Enhanced Attack on Reduced-Round AES

Mythos Preview developed a more efficient attack on a 7-round variant of AES-128 (the full version uses 10 rounds), improving the speed of previous best attacks by 200–800×.

The Möbius Bridge Technique

Using a meet-in-the-middle attack strategy, Mythos developed a new fingerprinting algorithm termed the "Möbius Bridge." This algorithm creates a fingerprint invariant to a specific guess that previous attacks had to enumerate (256 different values), thereby reducing the required work by a factor of 256. Despite the increased computational cost of the transform, Mythos implemented further optimizations to achieve the overall speedup.

Autonomous Discovery and Prompting

This result was achieved almost entirely autonomously via a scaffold that allowed the model to pose and test hypotheses. Notably, the model initially resisted the task, claiming that improving AES cryptanalysis was impossible. It only began producing novel results after the researchers used direct, non-technical prompting to encourage the model to act as a "top researcher" and avoid "low hanging fruit."

Additional Cryptographic Findings

Beyond HAWK and AES, Mythos Preview has identified other vulnerabilities in reduced-round or specific ciphers:

• Lightweight Encryption Algorithm (LEA): Developed a practical attack to recover a 13-round LEA key in under an hour on a modern desktop using fewer than 2²ⁿ encrypted plaintexts.
• Serpent-128: Identified a practical full key-recovery attack on 6 rounds of the 32-round Serpent-128 cipher.
• Minor Improvements: Achieved limited gains (<10×) on attacks against the Salsa20 stream cipher, the Poseidon hash function, and the SHA-1 hash function.

Implications for Cryptography and AI

These results suggest that the bottleneck in AI-driven research is shifting from the model's ability to generate hypotheses to the human ability to verify them. For the AES attack, while Mythos spent one week on the discovery, human researchers spent several hundred hours validating the claims.

Community Insights and Counterpoints

Discussion among the technical community highlights several critical perspectives on these findings:

• Theoretical vs. Unknown Weaknesses: Some critics argue that the HAWK finding is more of a "concrete mathematical counterexample" than a truly unknown weakness, as the theoretical possibility of the attack was already known.
• PQC Skepticism: The discovery of AI-driven attacks on post-quantum candidates has reinforced skepticism regarding the immediate transition to PQC, with some suggesting hybrid encryption (combining classical and PQC ciphers) as a safer alternative.
• Economic Cost: The $100,000 API cost per discovery has been noted as a significant investment, leading some to question if the same resources spent on human researchers would yield more reproducible long-term results.

To facilitate further study, Anthropic has partnered with academics from ETH Zurich, Tel Aviv University, and the University of Haifa to create CryptanalysisBench, a benchmark designed to evaluate the cryptanalytic capabilities of LLMs.