Anthropic released two new cryptography results this week. The company says its Claude Mythos Preview model played a lead role in both.
One result targets HAWK-256, a smaller test version of a post-quantum signature scheme called HAWK. The other improves an existing attack method on seven rounds of AES-128, a widely used encryption standard.
Anthropic published technical papers and code for both results. The company says the goal was to test how far AI can go in cryptanalysis, the practice of finding weaknesses in encryption systems.
How The HAWK Attack Works
HAWK is one of nine signature schemes still being reviewed by the National Institute of Standards and Technology. NIST advanced it to the third round of its post-quantum standards process in May 2026.
HAWK-256 is not a real security parameter. It is a challenge version meant to test researchers, separate from the stronger HAWK-512 and HAWK-1024 versions used in practice.
Mythos Preview found a lattice symmetry that earlier researchers had not used. That symmetry let the model build what Anthropic calls a shortcut path to recover a working secret key.
The attack does not recover the original secret key file. Instead, it produces a different key that can still sign messages the same way.
Anthropic says the full attack ran in about three hours and 42 minutes on a 96-core server. The company estimates the work needed to break HAWK-256 dropped sharply as a result.
HAWK-512 and HAWK-1024 remain out of reach. Anthropic says the gate count needed to attack those versions is still far too high for any real attack today.
AES Test Gets A Speed Boost
The second result involves AES-128, but only a reduced seven-round version, not the full ten-round cipher used in real systems.
Researchers already had a way to attack seven-round AES-128. Mythos Preview found a new mathematical shortcut, which Anthropic calls the Möbius Bridge, that removes one costly step from that method.
The change makes the attack 200 to 800 times faster, depending on how it is measured. It still requires an unrealistic number of chosen plaintexts, something no real attacker could easily obtain.
Anthropic says the model first refused to try, saying improving on AES was not possible. Researchers pushed it with follow-up prompts until it kept searching and eventually found the shortcut.
The model took about three days to find the initial idea. It generated close to one billion output tokens while refining the method over the following days.
Anthropic put the direct computing cost at around $100,000 for each project. Human researchers then spent hundreds of hours checking the work, with the AES verification alone taking close to a month.
Anthropic says the results follow last week's release of CryptanalysisBench, a 191-task test built with researchers from ETH Zurich, the University of Haifa, and other institutions. Several AI models broke a majority of its easier tasks during that test.
As of July 29, NIST still lists HAWK as a third-round candidate. Anthropic says it has not yet learned whether NIST or the HAWK team plan to change the scheme in response to the new findings.
