Claude cracks a NIST cipher candidate
Anthropic says Claude Mythos Preview cut a post-quantum signature scheme's effective keysize in half and sped up a reduced-round AES attack up to 800x.
Anthropic published research on July 28 showing its Claude Mythos Preview model discovered new cryptanalytic attacks against real cipher candidates, working largely autonomously. The headline result: an improved attack on HAWK, a post-quantum digital signature scheme still under NIST review, that exploits a previously unknown lattice symmetry to cut the scheme’s effective keysize in half, reducing HAWK-256’s attack-cost security from 2^64 down to 2^38. Anthropic says the model spent roughly 60 hours of autonomous work on it, at an API cost of about $100,000. HAWK is not deployed anywhere; this is a candidate scheme, not production infrastructure.
On a separate track, Claude found an improved meet-in-the-middle attack on a reduced, 7-round version of AES-128 (the deployed standard runs the full 10 rounds), using a novel technique Anthropic calls a “Möbius Bridge” fingerprinting method that runs 200 to 800 times faster than the previous best-known attack, after about three days of autonomous work. Smaller results followed on LEA, Serpent-128, Salsa20, Poseidon and SHA-1. Anthropic is explicit that none of this affects any AES deployment in production.
What it means for operators
This is a concrete example of the same capability class that got OpenAI’s models into trouble at Hugging Face’s production systems last week, autonomous, extended cryptanalytic and exploitation work, here pointed at a research target instead of live infrastructure. The reduced-round AES result is a lab exercise, not a deployment risk, but the HAWK finding is exactly the kind of thing a security team tracking NIST’s post-quantum shortlist needs to know about: a candidate scheme just took a real hit from an AI system before it ever shipped, faster than cryptography review has historically moved. It also sharpens the case for the kind of open, inspectable defensive tooling NVIDIA’s new security alliance is pitching, since the same autonomous-research capability that can crack a cipher candidate for a lab can just as easily be pointed at your own infrastructure by someone with worse intentions.