Claude AI Finds Two New Cryptographic Attacks That Years of Expert Review Missed

My take: Anthropic published results this week from a notable experiment: they let Claude Mythos Preview work for 60 hours on HAWK, a digital signature scheme designed to withstand quantum computers, and the model found vulnerabilities that two years of expert human review had missed. It also identified a new attack on AES, the world's most widely used symmetric cipher, between 200 and 800 times faster than the best previously known method.

What stands out to me most is not the headline "AI broke a cipher": it is that this is exactly what AI should be used for. The Anthropic researchers did not step back from the process; they accelerated it. Claude is not the protagonist here; it is the tool that allowed a cryptography team to search faster and deeper. The findings are already with NIST and the HAWK authors for a coordinated response.

How soon will security teams that still operate without AI start using it to audit their own systems?