What changed
CryptanalysisBench packages 191 tasks across six cryptographic primitive families and three difficulty tiers. Five frontier models solved 65–86% of known-break tasks and reportedly found previously unknown issues in SpoC and KINDI.
Why it matters
Cryptanalysis is a rare reasoning domain where candidate outputs can often be checked mechanically, making it useful for measuring both capability and emerging security risk.
What to do
Use the benchmark to stress-test candidate schemes in an isolated research environment. Require executable attacks and independent cryptographer review before escalating any finding.
Caveat
This is a non-peer-reviewed preprint from researchers including Anthropic employees. Novel findings and headline model comparisons still need independent reproduction.