First note from the UC Berkeley Agentic AI Summit:

Anthropic disclosed something worth being precise about, because the honest version is more interesting than the hyped one.

Claude Mythos (an agentic Claude system), working semi-autonomously with occasional human direction spent roughly 60 hours on HAWK, a post-quantum signature candidate still under NIST evaluation, not a deployed standard. It found a previously unexploited lattice symmetry that effectively halves HAWK's key strength. Real math, real finding, but the resulting attack is still exponential, and HAWK isn't protecting anything in production today.

Separately, a different Claude-driven scaffold worked almost entirely autonomously for three days against seven-round AES (not the ten-round AES-128 that actually protects your data) and produced an attack 200 to 800 times faster than the prior best-known method. That sounds alarming until you hit the actual requirement: the attack still needs on the order of 2^105 chosen plaintexts. That's not a number you get close to in practice. It's a stronger academic result, not a working exploit.

This isn't AI finding a bug in someone's implementation, which is the usual "AI found a security flaw" story. This is AI contributing directly to attacks on the underlying mathematics, the theory itself, not someone's sloppy code. And the more striking detail isn't the speedup. It's that humans became the bottleneck. The system generated the results faster than anyone could verify whether they were real.