Claude Reduces HAWK Key Strength by 94% in 60 Hours

Anthropic's Claude Mythos Preview weakened HAWK-256 security from 2^64 to 2^38 bits in 60 hours. The model also accelerated 7-round AES attacks 200-800x, highlighting AI-driven cryptanalysis risks.

Woofun AI reports that Anthropic’s Claude Mythos Preview model identified novel attack vectors against the NIST post-quantum signature candidate HAWK, reducing its effective key strength from 2^64 to 2^38 within 60 hours. The model independently conducted literature reviews and mathematical reasoning to achieve a 200 to 800 times speed increase in attacking 7-round AES, generating approximately 1 billion tokens before proposing core innovations. Anthropic noted that AI-generated cryptanalysis results now outpace human verification capabilities, prompting the launch of the CryptanalysisBench benchmark and coordinated disclosures with NIST authors. While HAWK remains undeployed and no production systems are currently affected, Anthropic warns that advancing AI capabilities may eventually enable practical attacks on deployed cryptographic infrastructure.