Claude Mythos just did what two years of expert cryptographers couldn’t.
Anthropic’s unreleased model found a mathematical weakness in HAWK — a leading post-quantum digital signature scheme that was still in NIST’s final selection round. In roughly 60 hours of work (and ~$100k in compute), it cut the effective key strength in half.
For the smallest parameter set, the cost of a full key recovery dropped from 264 2^{64} 264 operations to 238 2^{38} 238. That’s not a theoretical footnote. That’s a practical break on the challenge version. The developer has already withdrawn HAWK from the competition.
The same model also invented a new attack technique (“Möbius Bridge”) that speeds up the best known attack on 7-round AES by 200–800x. Full AES remains untouched. No production systems are affected today.
But the signal is clear: AI is now doing original cryptanalysis at a level that forces standards bodies to react in days, not years.
Post-quantum cryptography was supposed to be the long-term shield for everything from banking to blockchain. One of its most promising candidates just got sidelined by a language model.
The future of secure systems isn’t just about better math. It’s about whether that math can survive machines that never sleep and never miss a symmetry.
Nothing is broken yet. But the timeline just got a lot shorter.
Anthropic’s unreleased model found a mathematical weakness in HAWK — a leading post-quantum digital signature scheme that was still in NIST’s final selection round. In roughly 60 hours of work (and ~$100k in compute), it cut the effective key strength in half.
For the smallest parameter set, the cost of a full key recovery dropped from 264 2^{64} 264 operations to 238 2^{38} 238. That’s not a theoretical footnote. That’s a practical break on the challenge version. The developer has already withdrawn HAWK from the competition.
The same model also invented a new attack technique (“Möbius Bridge”) that speeds up the best known attack on 7-round AES by 200–800x. Full AES remains untouched. No production systems are affected today.
But the signal is clear: AI is now doing original cryptanalysis at a level that forces standards bodies to react in days, not years.
Post-quantum cryptography was supposed to be the long-term shield for everything from banking to blockchain. One of its most promising candidates just got sidelined by a language model.
The future of secure systems isn’t just about better math. It’s about whether that math can survive machines that never sleep and never miss a symmetry.
Nothing is broken yet. But the timeline just got a lot shorter.
