← Back to feed

Anthropic published two cryptanalysis papers on July 29, both outputs of Claude Mythos. The results open a new research discipline — AI-assisted cryptanalysis — and the more significant of the two directly affects a candidate post-quantum encryption standard.

The HAWK Result

HAWK is a proposed post-quantum signature scheme based on the module Lattice Isomorphism Problem (module-LIP). It was a reasonably advanced NIST evaluation candidate, positioned as a more efficient alternative to the Falcon signature scheme, which is already in the standardization process.

Mythos found a new key recovery algorithm against HAWK that roughly halves the scheme’s effective security level — in practical terms, this means HAWK would need to double its key sizes to compensate. Since HAWK’s entire justification is being more efficient than alternatives, the result makes the algorithm’s continued development very hard to defend. The attack does not transfer to Falcon, which is based on a different hard problem.

The attack is still exponential-time and is not practical against real deployed parameters. Anthropic provided a weakened “challenge instance” that Mythos’s code runs against in a few hours of wall-clock time. The result demonstrates the cryptanalytic weakness clearly, even if it doesn’t threaten live systems today.

Anthropic disclosed the finding to HAWK’s authors and to NIST before publication.

Matthew Green, a Johns Hopkins cryptography professor, noted that Mythos didn’t invent new mathematics: “What makes this genuinely interesting — and, frankly, a little embarrassing for the field — is that none of the ingredients are exotic.” The attack applied known tools more thoroughly. That framing matters: this is precisely the task where an AI with broad technical recall and no cognitive fatigue holds a structural edge over individual researchers.

The AES Result

The second paper improves an attack on 7-round AES by a modest constant factor. Full AES uses 10–14 rounds depending on key size. Published attacks against reduced-round AES go back to 2012–2013, and this result extends that prior work.

The improved attack requires 2^89 cipher operations and 2^105 chosen-plaintext encryptions. Neither figure is remotely practical. Standard AES is not threatened.

CryptanalysisBench

Anthropic released CryptanalysisBench alongside both papers — a new evaluation that measures AI systems’ ability to find cryptographic weaknesses. Mythos Preview’s full chain of thought during the AES analysis is also published.

The benchmark positions Anthropic at the center of a nascent research category. The HAWK result demonstrates it isn’t purely theoretical: AI cryptanalysis found a meaningful weakness in a real candidate standard. Whether the methodology scales to more structurally complex problems — full-round ciphers, deployed lattice schemes — is now an open and actively testable question.