Arcs Directions
Back to Timeline

Event record

Claude Mythos improves attacks on a post-quantum candidate and reduced-round AES

Anthropic reports that Claude Mythos helped improve a key-recovery attack against the non-deployed HAWK post-quantum signature candidate and an attack on seven-round AES-128.

Event Summary

Anthropic published cryptanalysis results on July 28, 2026, describing an improved attack against HAWK and an improved attack against a reduced-round AES variant. The HAWK result is relevant to a candidate considered in post-quantum standardization; the AES result targets seven of ten rounds and does not break full AES. After the disclosure, HAWK’s submission team withdrew the candidate from NIST’s additional digital-signatures process.

Context & Narrative

Using Claude Mythos Preview in an agentic research harness, Anthropic describes two main results. The first improves a key-recovery attack on HAWK, a proposed post-quantum signature scheme that had survived earlier expert review but was not deployed or standardized. Anthropic says the attack roughly halves the effective security level for the affected parameter set, while remaining exponential. The second result improves a meet-in-the-middle attack on seven-round AES-128 by 200–800 times under a chosen-plaintext research model. Independent cryptography analysis considers the HAWK result meaningful for the candidate's standardization prospects, but describes the AES result as incremental and impractical against production AES. The most durable signal is the model's ability to synthesize and extend existing cryptanalysis, together with the continuing need for human verification. After publication, the HAWK team confirmed that straightforward mitigation would make the candidate uncompetitive and withdrew it from NIST’s additional-signatures process; this did not change the event’s production-cryptography limitation.

Key Findings

  • Fact Grade B

    Anthropic reported an improved attack on the proposed HAWK signature scheme and an improved attack on a seven-round AES-128 variant using Claude Mythos Preview.

    Sources [1][2][3]
  • Impact Grade B

    The HAWK result may affect the candidate's post-quantum standardization prospects, while the broader capability signal is AI-assisted extension of cryptanalysis research.

    Sources [1][4]
  • Interpretation Grade B

    Independent analysis treats HAWK as the more consequential result and the AES result as a smaller technical increment rather than a practical break.

    Sources [1][4]
  • Limitation Grade C

    HAWK is not deployed, the attack remains exponential and was demonstrated on a challenge instance; the AES result targets seven rounds and requires an impractical chosen-plaintext setting, so neither result breaks production cryptography.

    Sources [1][4]
  • Fact Grade A

    After Anthropic’s disclosure, the HAWK submission team withdrew HAWK from NIST’s additional digital-signatures standardization process; NIST updated its Round 3 candidate page accordingly.

    Sources [5][6][7]

Impact Assessment

  • Capability Leap +2 · Medium-term

    The results show a frontier model extending published cryptanalysis into executable research workflows on difficult mathematical objects.

    Affected Groups: cryptographers, AI researchers, post-quantum standards bodies, security engineers

  • Paradigm Shift +2 · Long-term

    Cryptanalysis becomes a concrete example of AI-assisted research where model-generated hypotheses can extend expert literature, but must still be checked by specialists.

    Affected Groups: academic researchers, security standards organizations, AI safety teams, software maintainers

  • Risk Creation -1 · Medium-term

    More capable automated cryptanalysis could increase pressure on the review and migration pipeline if future findings affect deployed systems.

    Affected Groups: cryptographic users, standards bodies, software maintainers

Consensus & Sources

Significance L2
Category Capability Breakthrough / Safety & Ethics
Consensus Actively Debated
Impact Index 6/10