Anthropic let Claude Mythos Preview loose on cryptographic algorithms. It autonomously searched papers, formed hypotheses, wrote code, and ran experiments to find mathematical weaknesses. The model proposed two new attacks: one against HAWK, a post-quantum signature scheme not yet deployed, and another against a simplified version of AES. The HAWK attack took about 60 hours; the AES core idea came from the model itself.
HAWK is designed to "stamp" digital files and resist future quantum computers. Claude broke a test version called HAWK-256 and lowered the security estimate for larger parameters. The value is in finding problems with candidate algorithms early, not breaking live systems.
AES, by contrast, is the encryption standard used everywhere—websites, banks, you name it. But the full AES wasn't cracked. The experiment only targeted a 7-round version (three rounds stripped out) and still required about 2^105 specially crafted inputs. The claimed 200–800x speedup just made a theoretical attack that was already impractical slightly less impractical.
Claude can participate in cryptography research but can't do it alone. It found the key insight for the AES attack in a few days, but Anthropic researchers spent hundreds of hours and nearly a month verifying it. Right now, Claude speeds up research, doesn't replace experts, and certainly doesn't break real-world systems.