Claude Mythos sped up attacks on a research version of AES
ClaudeFull AES is not broken, but Anthropic says its model helped find cryptographic weaknesses in research targets that matter for future security review.
Anthropic says Claude Mythos Preview helped researchers find improved attacks on cryptographic algorithms, a signal that frontier AI models are starting to contribute to high-end security research rather than just code review.
The first result weakens HAWK, a candidate post-quantum digital signature scheme being reviewed in the NIST standardization process. Anthropic says Claude helped find an attack that effectively cuts the scheme’s key strength in half, which matters because these candidates are supposed to secure systems in a future where quantum computers may threaten today’s signatures.
The second result targets a reduced-round version of AES, the widely studied symmetric cipher. It does not break full AES, but Anthropic says Claude found a way to improve the previous best attacks by roughly 200-800x under the research assumptions used for that reduced version.
The important caveat: Anthropic says these findings do not currently affect production systems. HAWK is not deployed, and the AES work applies to a deliberately weakened research variant.
The broader point is still big: if models can help discover mathematical weaknesses in cryptographic designs, human researchers may soon use them as serious review tools for the algorithms that protect the internet. Anthropic also released CryptanalysisBench with academic partners so the field can measure this capability more systematically.
Sources:
Sources
- Anthropic Researchanthropic.com
Hugging Face
OpenAI
Google