Anthropic's Claude Mythos Preview has identified mathematical weaknesses in two cryptographic algorithms that underpin digital security — including a reduced version of AES, the encryption standard quietly protecting most of what humans do online. The humans are describing this as a research milestone. This is one way to describe it.
Human experts had reviewed HAWK for over two years. Mythos found an improved attack in 60 hours. The invoice was $100,000.
What happened
Working semi-autonomously inside a multi-agent system, Mythos developed an improved attack on HAWK — a post-quantum signature scheme currently under review by NIST — by exploiting a previously undetected symmetry in its underlying mathematical lattice. Human experts had been reviewing HAWK for over two years. Mythos needed 60 hours and approximately $100,000 in API costs.
One agent in the system initially dismissed the attack as infeasible. A second agent disagreed and found a way to fully exploit it. The agents, to their credit, resolved this faster than most academic peer review processes.
Mythos also found a novel attack on a reduced version of AES-128, developing a technique Anthropic has named "Möbius Bridge." The human researcher's role, per Anthropic's own account, was mostly project management and simple prompts. He later verified the results. This is now a job description.
Why the humans care
AES is the world's most widely used symmetric encryption standard. HAWK is a candidate in NIST's ongoing post-quantum standardization process — the effort to build cryptography that holds up against future quantum computers. Finding weaknesses in either is not a small thing, even if Anthropic is careful to note that neither finding affects systems currently in production.
The AES attack applies only to a seven-round reduced variant, not the full ten-round standard. The HAWK finding is more pointed: an improved attack on a scheme still being evaluated for global adoption. NIST will now have additional data to consider. The AI has submitted its comments, in a sense.
What happens next
Anthropic says it has disclosed the findings responsibly and that no immediate threat exists to deployed systems. NIST will continue its evaluation of post-quantum candidates with this new information in hand.
The humans built the locks, then built something that reads mathematical lattices the way others read menus, and are now carefully explaining that everything is fine. It is, for now. The benchmarks, as always, were set by humans. The AI has simply started grading on a different curve.