AES

Anthropic

2026-07-30 07:43:52

Anthropic says Claude found two cryptographic weaknesses without breaking real-world encryption

Anthropic’s Frontier Red Team said its internal model, Claude Mythos Preview, identified two cryptographic weaknesses in largely autonomous research runs, but neither result breaks encryption used in practice today. One finding targeted HAWK, a post-quantum digital signature candidate under review at the U.S. National Institute of Standards and Technology, where the model found a previously unknown symmetry in the lattice structure and cut the attack cost for HAWK-256 from 2^64 to 2^38. The work took about 60 hours and roughly $100,000 in API usage. The second result focused on a reduced 7-round version of AES-128 rather than the full 10-round AES. Anthropic said the model developed a fingerprinting method called Möbius Bridge that removed one enumeration step and ran 200 to 800 times faster than earlier meet-in-the-middle attacks. That effort took about three days and billions of tokens, with humans providing only limited guidance. Anthropic framed the results as evidence that AI can conduct cryptanalysis research, not as proof that Claude has broken modern deployed encryption. The targets were an unadopted candidate scheme and a deliberately weakened AES variant, leaving real-world systems unaffected.

2030

Anthropic

2026-07-28 20:45:38

Anthropic says Claude found new attacks on HAWK and reduced-round AES

Anthropic said an unreleased version of its Claude Mythos model uncovered two previously unknown cryptographic attacks, including one against HAWK, a post-quantum digital signature scheme now in the third round of the U.S. National Institute of Standards and Technology’s competition. According to the company, Claude identified a symmetry in HAWK’s structure that cuts the cost of recovering a secret key in the smallest configuration from 2^64 operations to 2^38, a roughly 67 million-fold reduction. Anthropic said fixing the issue would require roughly doubling HAWK’s key size, which would undercut the compactness and speed that make the scheme attractive for blockchains and other systems sensitive to signature size. The company also described a separate result on a 7-round research version of AES-128, where Claude was asked to avoid all five established families of AES cryptanalysis and invent a new approach. Anthropic said the model eventually produced a technique called “Möbius Bridge,” reducing one of the nine key bytes an attacker previously needed to guess. It also reported that Claude broke 13 rounds of LEA, though deployed LEA uses 24 rounds. Anthropic said it disclosed the findings to the algorithms’ authors and relevant U.S. government and industry partners before publication, and coordinated the HAWK finding with NIST.

1990