Claude Mythos Preview Breaks a Quantum-Resistant Cipher in 60 Hours for $100K
Anthropic's Mythos Preview autonomously cracked HAWK and accelerated an AES attack, marking the clearest public proof that frontier AI can do original cryptanalysis.
1. Claude Mythos Preview Breaks a Quantum-Resistant Cipher in 60 Hours for $100K
On July 28, 2026, Anthropic published research showing that Claude Mythos Preview autonomously found previously unknown weaknesses in two cryptographic systems. First: HAWK, a post-quantum digital signature scheme that had survived two years of expert review. Mythos Preview found an attack that cut HAWK's key strength in half, in 60 hours. Second: a reduced version of AES, one of the most scrutinized encryption algorithms in history. Within a week, Mythos Preview found a way to accelerate an existing attack on that version by 200 to 800 times. Each result cost roughly $100,000 in API usage. Anthropic disclosed findings in advance to the algorithms' authors and to US government and industry partners.
Neither result breaks systems in production today. HAWK has not been deployed, and the AES attack targets a weakened variant, not the full cipher. But the strategic signal is not about what broke. It is about what the result proves: a frontier model, running mostly autonomously with occasional human guidance, matched or exceeded the output of specialized human cryptanalysts on a hard, original research task. For NIST and the broader post-quantum standardization community, this changes the threat model. Algorithm review cycles measured in years now face an adversary that can run equivalent analysis in days. Organizations like Google, which has already deployed post-quantum cryptography in Chrome and Android, will need to treat AI-assisted cryptanalysis as part of their ongoing threat assessment, not a future concern.
Anthropic also released CryptanalysisBench, built with academics at ETH Zurich, Tel Aviv University, and the University of Haifa, to track LLM cryptanalysis capabilities over time. That benchmark matters as much as the individual results. It means the field now has a standardized way to measure how fast this capability is advancing. Watch for competing labs to run their own models against CryptanalysisBench, and for government partners to start asking what a $100,000 API budget buys an adversary next year.
Source: Anthropic on X