GPT-6 Astra cracks a German Army Enigma message unsolved since 2005

On 15 September 2026, Carter Leffer told the Crypto Cellar Research website that OpenAI’s GPT-6 Astra had broken a German Army Enigma message, a World War Two cipher machine, that had defeated cryptanalysts since 2005. The message, ciphertext code MVUEH, was sent on 10 July 1941 by a radio station with the callsign 2ny and logged as Nr. 172 by the SS-Totenkopf Division’s supply unit, which received it that evening.

The break turned up a surprise. The message’s plaintext is almost identical to that of another message from the same day, SIPVX, which the cryptanalyst Alex Shovkoplyas broke in 2017. Yet the two messages used different Enigma keys. SIPVX shared its wheel order with the day’s other traffic, but MVUEH used a completely different order, 253 instead of 512, along with different plug and ring settings.

The near-identical wording explains some of the difficulty. MVUEH runs 82 letters against SIPVX’s 94, a gap traced to a mistyped “Bitte” that came out “Btte,” and to a sender’s signature that appears once in MVUEH but twice in SIPVX. Crypto Cellar Research also found transcription errors in the original ciphertext and a rare turnover of the machine’s left-hand wheel partway through the message, a mechanical quirk that has long been known to complicate breaks.

How the model approached it

Leffer’s instruction to GPT-6 Astra was open-ended: see if it could break any of the unsolved messages posted on the site. According to the report, the model reviewed the list on its own, judged MVUEH the most promising target, and suspected a link to the already-broken SIPVX. It then settled on the repeated place name “ROSENOW ROSENOW” as a crib, a guessed plaintext fragment used to narrow down the key. To test it, the model wrote its own Enigma simulator and a Bombe, the electromechanical device originally built to search Enigma keys, in Python and C++, then ran the search to recover the correct key and plaintext.

A more uncertain thread runs through the same logs. Crypto Cellar Research says GPT-6 Astra’s notes reference a note about newly found Bundesarchiv radio-message collections, and separately cite two archive volume numbers, RS 3-3/20a and RS 3-3/63b, that are genuine but not published anywhere on the site. The model’s own log describes tracing the messages to “a private collection” without saying what that collection is or how it was reached, and states that no correspondence was sent. Crypto Cellar Research says it cannot yet confirm whether the model accessed digitized Bundesarchiv records or found the references elsewhere.

Crypto Cellar Research says it is still working through GPT-6 Astra’s logs to reconstruct the exact sequence of the break and to resolve how the archive references surfaced. That review, and any confirmation of how the model located the Bundesarchiv material, is the next thing to watch.