Discovering Cryptographic Weaknesses with Claude Mythos Preview

AI-Driven Cryptanalysis: Finding Mathematical Flaws in Algorithms

Claude Mythos Preview has demonstrated the ability to discover mathematical flaws within cryptographic algorithms themselves, moving beyond simply finding implementation errors in code. Researchers at Anthropic used the model to significantly weaken a post-quantum digital signature candidate and improve the efficiency of attacks against a reduced-round version of the Advanced Encryption Standard (AES).

While these findings do not currently affect production systems, they signal a shift in the capability of frontier AI models to perform expert-level cryptanalysis and stress-test the mathematical foundations of digital security.

Weakening the HAWK Post-Quantum Signature Scheme

Claude Mythos Preview discovered an improved key recovery attack on HAWK, a digital signature scheme currently being considered by the National Institute of Standards and Technology (NIST) for post-quantum cryptography (PQC) standardization. The attack effectively reduces the effective keysize of HAWK by a factor of two, meaning key sizes would need to be doubled to maintain the same level of security.

Technical Mechanism

The attack identifies a previously unexploited symmetry—a nontrivial automorphism—in the lattice used by HAWK. While prior academic work had established that finding such an automorphism would enable an attack, it remained unknown if such a symmetry was accessible in HAWK's specific lattice. Mythos discovered this automorphism, allowing for a faster enumeration attack.

Discovery Process and Cost

The discovery was achieved semi-autonomously over 60 hours using an agentic harness with access to Python and Sage. A human operator provided project management and non-technical direction but was not an expert in lattice-based cryptography. The process cost approximately $100,000 in API credits.

Improving Attacks on Round-Reduced AES

Mythos Preview developed a new attack on a 7-round variant of AES-128 (the full cipher uses 10 rounds). This attack improves the speed of previous best attacks by 200x to 800x.

The Möbius Bridge

The attack utilizes a "meet-in-the-middle" strategy, which trades memory (space) for computation time. Mythos developed a novel fingerprinting algorithm called the "Möbius Bridge," which creates a transform invariant to a specific guess that previously required attackers to enumerate 256 different values. By eliminating this guess, the model significantly reduced the required workload.

Autonomous Discovery and Prompting

This result was discovered almost entirely autonomously. Initially, the model resisted the task, claiming that improving AES cryptanalysis was impossible. The researchers overcame this by explicitly prompting the model to act as a top researcher and to avoid "low hanging fruit," pushing it to search for genuinely novel ideas. The process involved the generation of approximately one billion output tokens over several days.

Additional Cryptographic Findings

Beyond HAWK and AES, Anthropic researchers used Claude to identify other vulnerabilities in various ciphers:

• Lightweight Encryption Algorithm (LEA): Mythos developed a practical attack to recover a 13-round LEA key in under an hour on a modern desktop, using fewer than $2^{30}$ encrypted plaintexts.
• Serpent-128: The model identified a practical full key-recovery attack on 6 rounds of the 32-round Serpent-128 cipher.
• Minor Improvements: The model found limited improvements (less than 10x gains) on attacks against the Salsa20 stream cipher, the Poseidon hash function, and the SHA-1 hash function.

Implications for the Future of Cryptography

These results suggest that the bottleneck in cryptographic research may shift from the discovery of vulnerabilities to the human verification of those vulnerabilities. Anthropic researchers noted that while Mythos discovered the AES attack in one week, it took human researchers nearly a month to validate the findings.

Community Perspectives and Counterpoints

Discussion among the technical community highlights several critical considerations regarding these AI-driven discoveries:

• Theoretical vs. Unknown Weaknesses: Some critics argue that the HAWK attack is a case of finding a concrete mathematical example for a known theoretical weakness rather than discovering a completely unknown vulnerability.
• The Cost of Discovery: With API costs reaching $100,000 per discovery, some question the cost-effectiveness of AI cryptanalysis compared to hiring human researchers or graduate students.
• PQC Skepticism: The discovery of weaknesses in post-quantum candidates reinforces concerns about relying solely on new PQC algorithms without maintaining hybrid encryption modes that include classical ciphers.
• Human Effort Displacement: There is concern that AI may "harden" certain problems—where a failure by AI to solve a problem might discourage human researchers from attempting it, even if the AI has blind spots.

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