AI Breakthrough: LLMs Discover Novel Attacks on Cryptographic Schemes

In a groundbreaking development that has cybersecurity experts buzzing, a new benchmark called CryptanalysisBench reveals that cutting-edge Large Language Models (LLMs) are not just understanding complex code but actively finding new ways to break it. Anthropic frontier models, including Claude Mythos 5, alongside GPT-5.5 and GLM-5.2, didn't just reproduce old attacks; they uncovered novel vulnerabilities, like a key-recovery attack on the SpoC AEAD and a flaw in KINDI CCA-security proof. This marks a significant leap in AI's ability to perform mathematical cryptanalysis, pushing beyond theoretical understanding into practical exploit discovery. This isn't just an academic exercise. The CryptanalysisBench, comprising 191 tasks across six families of cryptographic primitives from NIST standardization competitions, acts as a rigorous testing ground. The ability of these LLMs to tackle and even break Tier 1 and Tier 2 schemes, which include cryptographic systems being considered for future digital security, signals a paradigm shift. Anthropic Claude Mythos Preview also exposed weaknesses in HAWK, a post-quantum digital signature scheme, and a reduced-round version of AES, demonstrating that AI is rapidly becoming a formidable force in analyzing the very foundations of our digital defenses. The immediate takeaway is clear: while these specific AI-found attacks don't yet threaten widely used production systems, they highlight the dual-use nature of advanced AI. These models can accelerate vulnerability discovery, acting as powerful tools for both attackers and defenders. Cybersecurity teams and cryptographers now face the urgent task of integrating AI-driven cryptanalysis into their defense strategies, using tools like CryptanalysisBench to proactively stress-test new cryptographic schemes before they're deployed. The race between AI finding weaknesses and humans patching them just got a whole lot faster.