Anthropic published a new Frontier Red Team report on July 28 describing how Claude Mythos Preview helped improve two cryptographic attacks: one against HAWK, a post-quantum signature candidate, and another against reduced-round AES.
HAWK is the sharper result
HAWK is still a candidate in NIST's post-quantum digital signature process. Anthropic says Mythos improved the best-known attack in 60 hours, effectively cutting its key strength in half. In the technical paper, the expected full key-recovery cost for the small HAWK-256 size moves from about 2^64 to 2^38.
This does not break all lattice cryptography, and it does not affect other NIST post-quantum signature candidates. It does suggest that AI-assisted review can become part of how cryptographic standards are stress-tested before deployment.
AES is a warning, not a break
The AES work is narrower. Full AES-128 has 10 rounds; Claude improved an attack on a 7-round reduced version used in academic cryptanalysis. Anthropic says the new Möbius Bridge fingerprint removes a 256-way guess and speeds up the previous best approach by 200 to 800 times, depending on the runtime measure.
The attack still assumes 2^105 chosen plaintexts and remains impractical. Anthropic explicitly says it does not break full AES and does not require production systems to change.
The bottleneck moves to verification
Humans remained in the loop. One researcher worked with Claude on HAWK, while another built a scaffold for the AES search. Anthropic says each main result cost roughly $100,000 in API usage, and the AES idea took a week to discover but nearly a month for researchers to gain confidence in its correctness.
The useful takeaway is not panic. It is that AI may compress the discovery phase of cryptographic research, leaving disclosure, validation and repair as the slower parts of the system.
Sources: Anthropic Frontier Red Team, Anthropic HAWK technical paper, Anthropic AES technical paper, NIST post-quantum signature project, CocoLoop; verified HAWK candidate status, 60-hour discovery, 2^64 to 2^38 key-recovery scope, AES-128 10-round versus 7-round scope, 200-800x speedup, 2^105 chosen plaintext assumption, roughly $100,000 API cost and production-system impact limits.