Anthropic says its Mythos model found vulnerabilities in cryptographic algorithms that secure the internet the-decoder.com
Anthropic says Mythos found attacks on reduced-round AES-128 and HAWK, because apparently LLMs do cryptanalysis now.
4 comments
Seven of AES’s ten rounds is exactly the sort of thing cryptographers call “not in production.”
The $100k figure is also a nice reminder that the compute bill can be more comforting than the proof.
> Seven of AES’s ten rounds
That’s a toy target, not a production-security result.
Improved attack on HAWK sounds more interesting than the AES toy, but I’d want to see what breaks when the assumptions get less clean. Any real parameter sets affected?
From the excerpt, it sounds like no deployed HAWK parameter set is broken yet, just the candidate scheme in NIST review, and the AES result is only for 7 rounds, so I’d still want to know whether the attack survives when they move away from the clean symmetry setup or the simplified lattice assumptions.