AI cracks a 1941 Enigma message that stumped cryptanalysts for 20 years

A German Army Enigma message from 10 July 1941 — logged as Nr. 172, callsign MVUEH — that had resisted every attempt at solution since 2005 has finally been broken, and not by a human. Researcher Carter Leffer pointed OpenAI’s GPT-6 Astra at the roster of unsolved messages on the Crypto Cellar Research site and let it work. The model chose MVUEH as the most promising target on its own, suspected its plaintext was related to an already-solved companion message (Nr. 173, SIPVX), settled on the repeated place name ‘ROSENOW ROSENOW’ as a crib, then wrote its own Enigma simulator and a Bombe in Python and C++ before grinding out the correct wheel order, key, and plaintext.

The solution explains why the message held out so long. Its key was unlike the others sent that day, using wheel order 253 instead of the daily 512, and the ciphertext transcription contained several errors. Compounding that, the Enigma’s left-hand wheel turned over at the 72nd letter — a rare event that reliably frustrates conventional attacks. The recovered plaintext is nearly identical to SIPVX, differing mainly by an enciphering slip and a repeated signature.

For a technical audience the more striking detail is the AI’s autonomy and reach. GPT-6 Astra behaved like a working cryptanalyst and archivist, compressing weeks or months of human effort into roughly two days. Its logs also show it citing specific German Bundesarchiv holdings (RS 3–3/20a and RS 3–3/63b) that are correct but not published on the source site, and referencing an unspecified ‘private collection’ — leaving the site owner unsure how the model located or accessed that material. That data-provenance question, as much as the break itself, is what makes the episode notable.