Anthropic reports its AI found flaws in cryptography humans studied for years

Anthropic says its Mythos model produced an improved attack on a post-quantum scheme in about 60 hours that experts had reviewed for over two years.

Anthropic says its AI model Claude Mythos Preview found mathematical weaknesses in cryptographic algorithms, including an improved attack on the post-quantum signature scheme HAWK and a new attack on a reduced version of AES. HAWK is a third-round candidate in NIST's competition for additional post-quantum signatures; human experts had reviewed it for more than two years, while Mythos produced its improved attack — exploiting a previously undetected symmetry in the underlying lattice — in about 60 hours. The AES work targeted a modified 7-of-10-round version and improved on prior attacks by a factor of 200 to 800 using a method the model called Möbius Bridge.

Each attack cost roughly $100,000 in API fees, and the AES run consumed around a billion tokens from only three substantive prompts. Human researchers, who were not cryptography experts, spent several hundred hours verifying the results. Anthropic shared its findings in advance with the US government and industry partners and coordinated disclosure with HAWK's authors.

For security, the promise is auditing at scale. If AI can surface subtle flaws in schemes meant to protect the internet — especially the post-quantum algorithms being chosen now to last decades — it could harden that infrastructure before weaknesses are standardized into everything. The same capability, of course, is a tool for whoever runs it.

The caveats are heavy. All claims come from Anthropic's own report; neither finding affects deployed systems; and both involve reduced or not-yet-standardized targets. The model initially refused both tasks before human encouragement, and human work was largely prompting and verification. Mythos Preview is not public.

Source: The Decoder