Anthropic's Claude Mythos cracks HAWK and weakened AES in under a week

Anthropic's Claude Mythos Preview, operating with a multi-agent scaffold, found two novel cryptographic attacks in under a week each, costing roughly $100,000 in API compute. The first attack targets HAWK, one of nine finalists in NIST's post-quantum cryptography standardization effort. The attack forces key sizes to double to maintain security, undermining HAWK's efficiency advantage as a lattice-based candidate. The discovery came from a human operator with no deep cryptographic expertise, with the key insight emerging from a pair of agents collaborating autonomously. The second attack improves on previous cryptanalysis of a weakened version of AES-128, reducing it from ten rounds to seven. Claude developed an algorithm called Möbius Bridge that makes the attack 200–800 times faster than prior work. This result was reached almost entirely autonomously: the model was asked to improve the best known attack, then designed and validated its own hypotheses over several days. A brief prompt exchange reminiscent of the recent ChatGPT Sol breakthrough helped spark the idea. Anthropic researchers then spent hundreds of hours validating the models' claims and preparing research papers.

Neither attack poses an immediate threat to production systems: the AES version is not deployed, and HAWK is not yet standardized. However, the demonstrations highlight the accelerating pace of AI-driven cryptanalysis. The multi-agent workflow and autonomous experimentation approach lowered the barrier for non-experts to produce cutting-edge research. Anthropic also released a redacted chain of thought from Mythos, offering rare insight into how frontier models reason about complex problems. The results suggest that as AI capability grows, cryptanalytic offense may outpace defense, a trend that security professionals must track closely.