CryptanalysisBench: Frontier Models Clear 65-86% of Already-Broken Ciphers but None Exceed 9% on Schemes With No Published Break
CryptanalysisBench / Anthropic·medium signal
Anthropic announced on 2026-07-29 the benchmark it built with ETH Zurich, Tel Aviv University, University of Haifa and TU Berlin -- 191 tasks across six primitive families drawn mainly from the AES, SHA-3, Lightweight Cryptography and Post-Quantum NIST competitions. Tier 1 covers 49 schemes with known practical breaks, where five frontier models scored 65-86% and Claude Mythos 5 led at 85.7%; Tier 2's 142 unbroken schemes held every model under 9% at full strength. The models did produce genuinely novel work: a key-recovery attack on the full SpoC AEAD and an error in KINDI's published CCA-security proof.