Anthropic's Claude AI discovers new weaknesses in post-quantum cryptography

Jul 28, 2026

Anthropic's Claude Mythos Preview AI has discovered significant new weaknesses in cryptographic systems. The AI found an attack on HAWK, a post-quantum candidate, reducing the cost to break it by 67 million times and forcing a redesign that negates its advantages. It also surpassed a 2013 record by accelerating an attack on a research version of AES, demonstrating a new AI-assisted workflow that may outpace human vulnerability verification.

HAWK cryptographic vulnerability

▪The attack on HAWK reduces the cost of recovering its smallest secret key from 2^64 operations to 2^38, a 67-million-fold reduction
▪Fixing the HAWK vulnerability requires doubling its key size, which undermines its primary advantages of compactness and speed
▪The Claude model also accelerated an attack on a 7-round research version of the AES cipher by 200 to 800 times
▪HAWK is a candidate in the third round of NIST's post-quantum signature competition

AES attack breakthrough

▪The AI developed a new technique for the AES attack called a "Möbius Bridge" after being barred from using established methods
▪The Claude Mythos Preview model also broke 13 rounds of LEA, a Korean national standard for lightweight encryption, in under an hour
▪The new AES attack surpasses a record set by human cryptographers in 2013

Claude Mythos Preview capabilities

▪The same model that found the crypto flaws also found 271 vulnerabilities in Firefox during internal testing
▪Each cryptographic discovery cost approximately $100,000 in API usage, and Anthropic staff spent several hundred hours verifying the AES work

AI-assisted cryptanalysis workflow

▪Anthropic warned that human verification processes may become a bottleneck, struggling to keep up with AI-discovered vulnerabilities
▪Anthropic created a benchmark called CryptanalysisBench with 191 cipher-breaking tasks drawn from NIST competitions
▪Anthropic's Claude AI discovered new weaknesses in two encryption systems, finding flaws missed by human experts
▪The HAWK research paper states that human contribution mainly consisted of directing, organizing, and verifying the AI's work