Claude Mythos cracks post quantum crypto and AES defenses

Anthropic’s Mythos result is one of the clearest demonstrations yet of frontier models acting as genuine research accelerators rather than just assistants. Cryptanalysis is a mature, technically demanding field where new attacks are rare and usually come from domain experts. The fact that an AI system, operating in an agentic scaffold with modest human guidance, can halve the effective key strength of a NIST‑track post‑quantum candidate and significantly sharpen a long‑studied AES variant attack signals that models are starting to contribute at the level of specialized PhD researchers.

For the race to AGI, this matters on two axes. First, it is a powerful proof‑of‑concept for “AI that builds better AI and better math,” shrinking iteration cycles in security‑critical domains. Second, it is an early warning about dual‑use risk: the same capabilities that help harden crypto standards could, in less careful hands, undermine existing systems. Anthropic’s emphasis that no deployed schemes are broken is reassuring, but the trajectory is clear. If models can autonomously explore large mathematical search spaces, we should expect them to uncover non‑obvious vulnerabilities across the digital stack.

Strategically, this pushes frontier labs further into a security‑policy spotlight and strengthens the argument that AGI‑class systems will emerge first inside organizations that pair massive compute with tightly integrated safety and applied‑research teams.