A large language model recovered the plaintext of a World War II Enigma message that had gone unbroken on the Crypto Cellar Research archive. Bruce Schneier published the account on his security blog, attributing it to the Crypto Cellar researchers themselves: they say Carter Leffer gave GPT-6 Astra one instruction, attempt any of the site’s unsolved ciphers, and left the rest to the model.
According to that account, the system picked message Nr. 172 (ciphertext MVUEH), linked it to a second unsolved message, and settled on “ROSENOW ROSENOW,” the repeated place name it used as a crib, a known-plaintext guess for attacking the cipher. It then built its own Python and C++ Enigma simulator and Bombe solver, using them to recover the correct key and plaintext.
“We are still analysing the GPT-6 Astra logs,” the researchers wrote, adding that they want to understand precisely what let the model succeed. Schneier has not independently verified that process; the account is the researchers’ own telling, not his reporting.
Cryptanalysis has tracked machine intelligence since Turing’s Bombe. A system that picks its own target and builds its own tools against a genuine unsolved historical cipher gives builders a concrete benchmark, beyond scripted leaderboards.
Reported by Bruce Schneier on his blog, Schneier on Security, citing the Crypto Cellar Research team, September 22, 2026.
