AI cracks AES and threatens crypto security for $100k
anthropic's mythos model discovered attacks on aes and hawk in under 60 hours for $100k, signaling that ai can now autonomously find critical flaws in cryptographic standards.
anthropic reports that its mythos preview model uncovered a mathematical weakness in a 7-round version of aes-128 and a new attack on the hawk lattice-based scheme. the model ran in a multi-agent system using a scaffold for hypothesis testing, costing about $100,000 in api compute, while humans provided only project management and sporadic prompting. the findings highlight a shift where ai can surpass human intuition in specialized fields, finding symmetries and attacks in hours that experts missed for years. this suggests ai could significantly accelerate the discovery of zero-day vulnerabilities in cryptography, necessitating a reevaluation of security assumptions for future standards.