HAWK, a digital signature algorithm vying to become a standard defense against quantum computer attacks, is dead. Its creator withdrew it from NIST’s post-quantum cryptography competition Tuesday, a day after Anthropic revealed its Mythos security model had found a flaw that broke the scheme. The algorithm had made it to the third round of NIST’s rigorous testing gauntlet — a phase specifically designed to catch this kind of deep mathematical weakness. It didn’t survive.
Anthropic didn’t stop at HAWK. In the same announcement, the company said Mythos also found weaknesses in the mathematical problems underpinning the widely used AES cipher. But before anyone panics about their bank account, a few crucial caveats apply. Anthropic tested weakened “challenge instances” of these cryptosystems, not the fully hardened production versions. That’s standard practice for adversarial peer review, but it means the real-world impact is blunted. The model didn’t shatter AES; it found methods to moderately reduce the work required to defeat these stripped-down versions.
Even with that progress, the underlying primitives — the fundamental math problems that form the building blocks of these systems — remain safe for now. This is an incremental step, not a catastrophic break. But it’s a meaningful one. The ability to automate the discovery of subtle, exploitable properties in cryptographic algorithms would fundamentally change the game, shifting some power from the defenders who design these systems to the attackers probing them. Anthropic’s model spotted a real flaw that human review had missed, sending a finalist packing.
How much of this is genuine breakthrough versus marketing muscle is the honest question. Anthropic is, after all, selling a product. But the result is a tangible pelt on the wall. A NIST candidate is gone. The finding suggests we’re entering an era where AI isn’t just a tool for security but an active participant in the mathematical arms race that defines cryptography, and that’s a development worth watching closely, skepticism intact.
