The era of manual cryptanalysis is hitting a computational wall. Anthropic has unveiled Claude Mythos Preview—a model that identified mathematical weaknesses in widely accepted encryption standards, including AES. While these vulnerabilities don't yet allow for real-time hijacking of bank accounts, the methodology is a total game-changer. For $100,000 in API bills, a machine accomplished what typically takes specialized experts years of their lives. This isn't the end of the internet; it is a collapse of the barrier to entry for the major leagues of cyber espionage.
The Autonomous Attack Vector
Anthropic deployed Mythos Preview within a multi-agent system where the AI was effectively left to its own devices. The model targeted HAWK—a post-quantum signature scheme that the U.S. National Institute of Standards and Technology (NIST) has been meticulously analyzing for the past two years. In just 60 hours, Mythos discovered a way to optimize an attack on the algorithm by identifying a hidden symmetry in the mathematical lattice. This was no "hallucination" or a lucky prompt: one agent initially deemed the task impossible, but a second found a loophole and drove the exploit to its logical conclusion.
"Mythos Preview found an improved attack in just 60 hours, whereas experts had been studying HAWK for over two years."
The human role in this process has degraded to that of a resource manager. Anthropic researchers admitted they lack deep expertise in lattice-based cryptography; they simply provided the inputs and verified the outputs. Deep expertise is no longer the bottleneck; success now hinges on token generation budgets and the ability to orchestrate computational flows.
Breaking the Resistance to AES
When attacking AES-128, the model initially displayed human-like stubbornness, claiming that existing methods could not be improved. Only after persistent demands to seek "fundamentally new ideas" did Mythos develop a fingerprinting method dubbed the Möbius Bridge. Applied to a 7-round version of AES, this technique proved to be 200–800 times more efficient than any previously known attack. The price tag for this breakthrough was that same $100,000.
"The method eliminates one of the guesses an attacker must make and outperforms the best known attacks by 200–800 times."
Anthropic’s data confirms a shift: hacking efficiency is now directly proportional to budget rather than individual talent. The company has already shared its findings with NIST and the U.S. government. The signal is clear: the intersection of AI and national security has moved from theoretical risk to measurable API consumption. The lifespan of current encryption standards has been drastically shortened. Where developing an attack once required hiring rare geniuses, state actors and major players can now iterate exploits at the speed of cloud computing. Migrating to quantum-resistant encryption is no longer a ten-year roadmap item; it is a race against time before the cost of automated vulnerability discovery drops by another order of magnitude.