Claude Mythos broke a weakened AES; 800 times faster than any human method Anthropic says its most powerful AI model has found new mathematical weaknesses in the encryption that holds the internet together. Claude Mythos Preview, a model still not available to the public, worked out an improved way to attack a stripped-down version of AES, the cipher sitting underneath online bank transfers, encrypted chats, Wi-Fi networks and stored data. The company published the findings on Tuesday alongside two research papers.Nothing you use today is broken. Mythos attacked a 7-round variant of AES-128, which normally runs 10 rounds, and its method is 200 to 800 times faster than the previous best. The second finding bites harder. Mythos halved the effective key strength of HAWK, a post-quantum signature scheme shortlisted by the US National Institute of Standards and Technology, after HAWK had already survived two years and two rounds of expert human review. It took the model 60 hours. Anthropic shared both results with US government and industry partners before publishing.
Why cracking a “reduced-round” cipher still matters to anyone with a banking app
Cryptographers routinely weaken algorithms on purpose to see where the cracks might appear later. Think of it as solving an easier version of a hard maths problem to spot the pattern. The AES attack also assumes an attacker can request 2¹⁰⁵ chosen encryptions, which is not happening on any machine that exists. AES was adopted in 2001 and is believed to have roughly as many possible keys as there are atoms in the observable universe.The point is the trajectory, not the damage. Anthropic researcher Nicholas Carlini, who previously worked at Google, said of last year’s models: “They could not do problems that I could do when I was 16.”
Anthropic’s model refused the job first, then burned a billion tokens finishing it
The discovery process is the strangest part of the story. Claude initially declined to work on AES because it was convinced no improvement was possible, telling researchers there was nothing easy left to find in the most-studied block cipher around. A researcher nudged it with a typo-riddled prompt noting that models assume the task is impossible and so never try. Claude then rewrote its own agent harness to hunt for novel ideas.Over the next three days it produced several hundred million tokens and landed on a technique it named the Möbius Bridge, a fingerprinting method that removes one of the 2⁵⁶ guesses an attacker previously had to make. By the end it had output about a billion tokens. Each of the two results cost roughly $100,000 in API spend. Two Anthropic researchers, neither a cryptography expert, then spent close to a month confirming the maths held up.
What the findings signal for NIST standards, the quantum race and AI safety rules
For HAWK, the practical fallout is immediate. The expected cost of recovering a HAWK-256 key drops from 2⁶⁴ to 2³⁸, and doubling key sizes to compensate strips away much of what made the scheme attractive.Anthropic also flagged a working attack on 13-round LEA that recovers a key in under an hour on a desktop, plus smaller gains against Serpent-128, Salsa20 and SHA-1. It has released CryptanalysisBench, built with ETH Zurich, Tel Aviv University and the University of Haifa, to track how fast this capability grows.Mythos remains restricted. When it launched in April, Anthropic limited access to select government agencies after the model proved unusually good at finding and exploiting software bugs. Glenn Gerstell, former general counsel at the NSA, questioned whether anyone should feel comfortable assuming strong encryption stays safe two years out.