Claude Mythos Preview has achieved what seemed to be beyond even the reach of the best cryptographers. Anthropic’s model has found ways to attack several encryption algorithms and has shown that artificial intelligence can take part in top-level mathematical research. An advance that confirms Mythos’s capabilities in cybersecurity and that, while it doesn’t pose a real risk yet, is raising concerns.
A much faster attack against AES
During the tests, as reported in the NYT, Claude analyzed a seven-round version of AES-128 —the full standard used on the internet uses ten— and developed a method between 200 and 800 times faster than the best known attack for recovering the data. Given that AES encryption protects web traffic, our wireless networks, storage, card payments, and most communications, the finding challenges one of the internet’s core technical foundations.
The result remains confined to research, for now, because the full version of AES remains secure and today’s banking and communications systems continue to be protected. These tests with simplified variants of the encryption algorithm help identify useful patterns for strengthening future standards, especially given the growing volume of vulnerabilities discovered through AI. The situation is still noteworthy, because what would have seemed like science fiction a few years ago, for example image generation, AI can now do in seconds.
Claude also takes on post-quantum encryption
Mythos also refined an attack on HAWK, a digital signature system that is a candidate to become the post-quantum standard. In about 60 hours it halved the effective strength of its keys and showed us once again how quickly AI can analyze even the most complex cryptographic designs.
Anthropic is steering these capabilities toward defense, as we’ve also seen with Claude Security, but what we take for granted today was, just a few months ago, simply unthinkable.
