Anthropic’s Claude Mythos Preview recovered signing keys from HAWK-256, a post-quantum encryption candidate under review by the US National Institute of Standards and Technology. The AI model also discovered a novel attack that speeds up seven-round AES-128 cracking by 200 to 800 times. Neither result threatens systems currently in production.
As per Anthropic:
When we launched Claude Mythos Preview, we showed it was able to autonomously find and exploit vulnerabilities in almost every piece of software we pointed it at. This included several major cryptographic libraries—shared collections of code that are used to encrypt data.
The HAWK result carries particular weight. HAWK is the only lattice-based scheme among nine candidates NIST advanced to its third-round post-quantum digital signature process in May 2026. Claude found a hidden symmetry in the lattice structure that prior researchers identified as theoretically exploitable but could not practically attack. The AI model exploited that symmetry to recover functionally equivalent signing material from HAWK-256 public keys.
The numbers tell the story. Claude reduced HAWK-256’s expected key-recovery work factor from 2^64 to 2^38. Anthropic’s released implementation completes full key recovery in approximately three hours and 42 minutes on a 96-core server. The attack only targets HAWK-256, a challenge parameter. Larger variants HAWK-512 and HAWK-1024 remain impractical to break.
The AES discovery is equally striking. Claude developed what Anthropic calls the “Mobius Bridge,” an invariant fingerprint that eliminates a 256-way guessing step from existing attacks on seven-round AES-128. This speeds up the attack dramatically. However, it still requires approximately 2^105 chosen plaintexts, placing it firmly outside real-world feasibility.
The process itself matters as much as the results. Mythos Preview conducted the HAWK research largely autonomously over 60 hours. The AES work took roughly three days and generated approximately one billion output tokens. Each project cost about $100,000 in API usage. Human researchers spent significantly longer verifying the results than the AI spent discovering them. Two researchers needed nearly a month to confirm the AES finding was correct.
Anthropic published both findings alongside technical papers, reproducibility code, and a NIST forum announcement. As of July 29, NIST still lists HAWK as a third-round candidate. No independent reproduction of the HAWK-256 recovery has been publicly confirmed.
According to them:
This is not the first time that language models have performed research-level mathematics. In just the last few months, researchers from Google have used Gemini to resolve several open Erdős problems, researchers from OpenAI have used GPT to resolve the unit distance conjecture (a particularly challenging Erdős problem), and earlier this month we announced that Claude Fable 5 resolved the Jacobian Conjecture. Our result here—that Claude is able to perform cryptographic research at the level of top experts—indicates that these same capabilities also have applications in the field of cryptography, and thus may soon have more practical consequences… The attacks described in these two papers are the strongest attacks we have found to date. We are sharing them after a period of consultation with US government and industry leaders. But as we develop increasingly powerful cryptanalytic results, it would be prudent to consider how researchers should react if a language model were to discover vulnerabilities in cryptosystems where attacks do have an immediate real-world impact.
The model initially refused to attempt the AES work. It insisted that improving on AES was impossible. Anthropic published the researcher’s blunt follow-up prompts that pushed Claude to keep investigating.
