Claude did not break AES. The signal still matters

Subject: Claude did not break AES. The signal still matters

Preview: Mythos improved attacks on HAWK and a deliberately reduced version of AES.

Claude Mythos Preview helped uncover new mathematical weaknesses in HAWK, a post-quantum signature scheme still under review, and in a deliberately reduced version of AES. The results do not break production systems in use today.

Anthropic says the model found a previously unknown attack on HAWK in about 60 hours, substantially reducing its security margin. On reduced-round AES, it reportedly made an existing attack 200 to 800 times faster. Each research effort cost roughly $100,000 in API usage.

Cryptanalysis combines mathematical reasoning, sustained exploration and hard verification. An agent contributing to new results in this field suggests that AI can accelerate the testing of the building blocks protecting digital communications and identity.

Full AES remains intact and HAWK is not deployed. The model received occasional human guidance, and experts still had to validate the work. This is a story about research speed, not the immediate collapse of encryption.

Cyber leaders should track capabilities as well as incidents. Use AI to review algorithms and cryptographic libraries, with independent verification, coordinated disclosure and strict access boundaries.

The near-term risk is not that all encryption breaks today, but that offensive research changes speed.