Anthropic Finds Cryptographic Algorithm Flaws with AI Model

Anthropic's Frontier Red Team published two research papers describing new attacks that its Claude Mythos Preview model found against cryptographic algorithms. In the first, the model identified a previously unknown mathematical symmetry in HAWK, a candidate in NIST's post-quantum cryptography process, reducing the scheme's effective security margin. The finding took roughly 60 hours of model work on a design that had already survived about two years of open review by human cryptographers.

The second result targets AES. The model devised an attack technique it named the Mobius Bridge, which makes an existing attack on a deliberately weakened research variant of AES between 200 and 800 times faster. Both results carry an important limit that is easy to lose in the headline: neither breaks anything protecting real-world data today. HAWK is not deployed in production anywhere, and the AES work applies to a reduced-round research variant, not the full cipher running in browsers, payment rails or encrypted storage.

What changed is the capability curve, not the threat level. Language models were unable to perform meaningful cryptanalysis of even elementary ciphers not long ago, and are now surfacing structural flaws in designs that withstood years of expert scrutiny. That has practical consequences for how standards bodies run review cycles, and it lands on a live commercial question for Anthropic, which has been positioning frontier-model capability in security research as a differentiator. Related coverage notes that the same capability curve accelerates the offensive side, which is the argument security vendors use for machine-led defense. The near-term marker is whether NIST's evaluation process adjusts in response and whether other labs reproduce the results.