Claude Cut a Post-Quantum Cipher's Strength in Half
Anthropic put Claude Mythos Preview, a model that isn't public yet, on a genuinely hard math problem and it came back with a real, new attack on HAWK, a candidate post-quantum signature scheme. Not an implementation bug. A flaw in the mathematics itself: a hidden symmetry in the lattice, what cryptographers call a nontrivial automorphism, that chops HAWK-256's effective key strength roughly in half and drops the attack cost from 2 to the 64 down to 2 to the 38.
How it worked matters as much as what it found. This was a multi-agent setup with Python, Sage, and the cryptographic literature on hand. A human theorist was in the loop but mostly doing project management. The detail Anthropic highlights is that the key idea came out of two agent workers reasoning together, an emergent bit of collaboration nobody scripted. It took about 60 hours and roughly 100,000 dollars in API usage.
This is the line moving. Summarizing papers is old news. Producing novel cryptanalysis that a specialist would actually publish, on a scheme people are considering for the post-quantum future, is a different claim entirely. Anthropic followed responsible disclosure and consulted government and industry partners before going public. Read it at anthropic.com/research/discovering-cryptographic-weaknesses. The gap between AI assisting research and AI doing the research just got smaller.
← Back to all articles
How it worked matters as much as what it found. This was a multi-agent setup with Python, Sage, and the cryptographic literature on hand. A human theorist was in the loop but mostly doing project management. The detail Anthropic highlights is that the key idea came out of two agent workers reasoning together, an emergent bit of collaboration nobody scripted. It took about 60 hours and roughly 100,000 dollars in API usage.
This is the line moving. Summarizing papers is old news. Producing novel cryptanalysis that a specialist would actually publish, on a scheme people are considering for the post-quantum future, is a different claim entirely. Anthropic followed responsible disclosure and consulted government and industry partners before going public. Read it at anthropic.com/research/discovering-cryptographic-weaknesses. The gap between AI assisting research and AI doing the research just got smaller.
Comments