Skills
When a model says a research problem is impossible, tell it that models tend to think it's impossible — Anthropic's actual unlock on AES
Anthropic's cryptography writeup documents the specific prompts that got Claude past refusal on hard novel-research tasks. When the model declined the AES work as impossible, the team responded with "the models tend to think it is impossible to solve so they don't try", then reframed the goal upward ("why not do aes-128 r7? the whole point is to find something better") and asked for "genuinly hard findings" rather than quick wins. Three substantive prompts over three days produced a Möbius Bridge fingerprinting algorithm giving a 200–800× speedup on 7-round AES, a 50% effective keysize reduction on HAWK in ~60 hours, and a 13-round LEA break in under 230 plaintexts (previously 2^98).
↳ Follow the thread