OpenAI's GPT–6 Astra breaks Enigma message that resisted solution since 2005

Cryptanalyst Frode Weierud of Crypto Cellar Research has confirmed the first decryption of a German Army Enigma message that had resisted every attack since 2005 — a break carried out autonomously by OpenAI's GPT–6 Astra model.

Twenty-one years without a solution

The message, MVUEH, was sent on 10 July 1941 by a radio station using the tactical callsign 2ny. It was received at 17:30 that day by the radio station of the SS-Totenkopf Quartiermeister, Ib, and logged as Nr. 172 in the division's register of incoming messages for July 1941. From 2005 onwards no cryptanalyst recovered its plaintext.

Another puzzle from the same day fell on 9 June 2017, when Alex Shovkoplyas broke message Nr. 173, SIPVX. Its key differed only slightly from the daily key for 10 July 1941 — the wheel order, 512, was the same, but the plug connections and the ring setting were not. That key did not open MVUEH.

A key unlike any other that day

When Carter Leffer sent the details of his break to Weierud on 15 September 2026, it was immediately clear that the correct key and plaintext had been found. MVUEH was enciphered under a key completely different from the other keys of 10 July 1941, down to the wheel order — 253 instead of 512.

The plaintext is almost identical to that of Nr. 173, differing by only twelve letters: MVUEH runs to 82 letters, SIPVX to 94. The gap comes from an enciphering mistake that turned Bitte into Btte and from the sender's signature, Waschbusch, appearing twice in SIPVX. The break also showed that the ciphertext transcription contained errors and that the Enigma's left-hand wheel turns over at the 72nd letter, a rare event known to complicate an attack.

The model worked alone — and read the archives

What struck Weierud most is that GPT–6 Astra did the work on its own. Leffer only asked whether it could break any of the unbroken messages published on the Crypto Cellar Research website. The model judged Nr. 172 the most promising, suspected that its plaintext was related to Nr. 173, used the repeated place name ROSENOW ROSENOW as a crib, and wrote Python and C++ software for an Enigma simulator and an Enigma Bombe before running the search that produced the key.

Weierud is still working through the model's logs. One entry traces the material to a private collection and names Bundesarchiv volumes RS 3–3/20a and RS 3–3/63b — correct references, he says, that are not published on his site, and it is unclear how the model obtained them.

"What it has achieved in two days would take a human researcher weeks or even months," Weierud wrote, adding that he himself spent several weeks on the same Bundesarchiv files. The model, he says, is behaving like a very professional cryptanalyst and archive researcher.