Anthropic Claude Mythos Preview discovers improved attacks on HAWK and reduced-round AES

TL;DR

Anthropic released two new cryptanalysis results generated largely autonomously by Claude Mythos Preview: an attack that halves the effective key size of the post‑quantum signature scheme HAWK, and a meet‑in‑the‑middle attack on 7‑round AES that is 200–800× faster than prior work. Both attacks are theoretical and do not affect deployed systems, but they illustrate how powerful language models can discover algorithmic flaws.

Overview of the research announcement

Anthropic’s blog post "Discovering cryptographic weaknesses with Claude" details how the Claude Mythos Preview model was used to find new attacks on two well‑studied cryptographic primitives. The work was carried out with minimal human guidance, cost roughly $100,000 in API usage per attack, and was disclosed responsibly to NIST and the original authors of HAWK. The findings are documented in two separate papers and are accompanied by open‑source demonstration code and a benchmark suite (CryptanalysisBench) for future LLM‑driven cryptanalysis.

Improved key‑recovery attack on HAWK

Core result

Claude Mythos Preview identified a nontrivial automorphism in the lattice underlying HAWK, enabling a faster enumeration attack that reduces the scheme’s effective key size by a factor of two. For the HAWK‑256 parameter set, the expected attack cost drops from (2^{64}) to (2^{38}).

Technical basis

  • HAWK’s security relies on the hardness of the Lattice Isomorphism Problem.
  • Prior work (e.g., ePrint 2025/928) proved that efficiently finding such an automorphism would permit an attack, but did not demonstrate its existence in HAWK’s lattice.
  • Mythos discovered the automorphism, built a verification pipeline in Sage/Python, and confirmed the attack end‑to‑end.

Discovery process

  • A Claude Code‑like multi‑agent harness coordinated several worker agents with access to computational tools and the cryptographic literature.
  • Human oversight was limited to project management; the key insight emerged from a dialogue between two agents, one of which initially dismissed the idea before the other refined it.
  • Total discovery time: ~60 hours; API cost: ~$100 k.

Impact assessment

  • The attack is exponential, not polynomial, and only affects HAWK, a NIST third‑round candidate that is not yet deployed.
  • It demonstrates that AI can uncover structural weaknesses that survived two years of expert review.
  • The result suggests that future NIST candidates should be evaluated with AI‑assisted analysis to avoid similar oversights.

Faster meet‑in‑the‑middle attack on 7‑round AES

Core result

Claude Mythos Preview introduced a "Möbius Bridge" fingerprinting algorithm that eliminates a 256‑fold enumeration step in the best known meet‑in‑the‑middle attack on 7‑round AES‑128, yielding a 200–800× speedup.

Technical basis

  • The attack assumes a chosen‑plaintext model with up to (2^{105}) queries, matching prior academic settings.
  • The Möbius Bridge creates a transform invariant to the 256 guesses required in earlier attacks, reducing the number of table lookups.
  • Additional optimizations offset the higher computational cost of the transform.

Discovery process

  • Researchers built a scaffold that let Claude pose hypotheses, run experiments, and iterate autonomously.
  • Initial prompts were required to overcome Claude’s self‑imposed belief that AES‑128 was "genuinely hard".
  • Over three days Claude generated hundreds of millions of tokens, eventually proposing the Möbius Bridge after roughly one billion tokens.
  • Human effort focused on validating the mathematics and preparing the paper; validation took several hundred hours.

Impact assessment

  • The attack targets a 7‑round reduced version of AES, not the full 10‑round cipher, so it has no immediate practical impact.
  • It extends a long line of reduced‑round AES research, showing that LLMs can accelerate the discovery of subtle algorithmic improvements.

Supporting infrastructure: CryptanalysisBench

Anthropic partnered with ETH Zurich, Tel Aviv University, and TU Berlin to release CryptanalysisBench (arXiv 2607.18538), a benchmark suite that packages multiple ciphers for systematic evaluation of LLM cryptanalytic abilities. The benchmark is intended to facilitate reproducibility and to track future model capabilities.

Additional preliminary findings

  • LEA‑13 attack – Mythos discovered a key‑recovery attack on 13‑round LEA requiring (<2^{30}) plaintexts and under an hour on a desktop. The full 24‑round LEA remains secure.
  • Serpent‑6 attack – A full‑key recovery method for 6‑round Serpent‑128 improves on prior work that needed (>2^{70}) plaintexts.
  • Minor gains (<10×) were observed against Salsa20, Poseidon, and SHA‑1.
  • These results are still being validated and will be released in future publications.

Implications for cryptography and security research

  • AI‑augmented cryptanalysis – The HAWK and AES results prove that frontier LLMs can perform research‑level cryptanalysis, potentially accelerating the discovery of weaknesses in both candidate and deployed schemes.
  • Human bottleneck – While models can generate attacks quickly, human experts still spend substantial time verifying correctness, especially for complex attacks like the AES Möbius Bridge.
  • Policy considerations – Anthropic followed responsible disclosure, informing NIST and HAWK authors before public release. As AI‑driven attacks become more powerful, coordinated disclosure frameworks will be essential.
  • Future workflow – The authors anticipate that AI will become a standard tool for adversarial review, helping designers iterate on more resilient algorithms and possibly contributing to the design of next‑generation cryptographic primitives.

Conclusion

Anthropic’s announcement demonstrates that Claude Mythos Preview can autonomously uncover mathematically non‑trivial attacks on both post‑quantum signatures and symmetric ciphers. Although the attacks target non‑deployed or reduced‑round versions, they highlight a paradigm shift: language models are now capable of contributing substantive cryptanalytic research, prompting the security community to adapt verification processes and to integrate AI tools into the standard cryptographic review pipeline.

Links to full papers

Sources

Related