AI · 3h ago
New Benchmark Shows LLMs Can Break Real Cryptography
Researchers introduced CryptanalysisBench, a benchmark testing LLMs' ability to find cryptographic attacks. Frontier models broke 65-86% of historically broken schemes and discovered novel attacks, including a key-recovery flaw in SpoC AEAD. The findings suggest AI is increasingly capable of automated cryptanalysis, with implications for digital security.
Meridian48 take
The benchmark is impressive, but real-world impact depends on whether these attacks scale to modern, production-grade encryption.
llm-cryptanalysisai-security