An 82-letter German Army signal from July 1941 resisted every attempt to break it for two decades. OpenAI’s model picked it as a target on its own, wrote its own bombe, and cracked it.
An Enigma message that had stayed unbroken since it was published online in 2005 has finally given up its plaintext, and the codebreaker was OpenAI’s GPT-6 Astra. The message, catalogued as MVUEH, was sent on 10 July 1941 and logged at 17:30 that day by the radio station of the SS-Totenkopf Division’s quartermaster section as incoming message Nr. 172. Frode Weierud, who runs the Crypto Cellar Research archive where the message was posted, validated the break and has published a write-up.
The content itself is mundane. Decrypted, it asks for the route of march, reports the sender’s position as Rosenow (twice), and requests an immediate radio reply. It is signed Waschbusch. What makes it interesting is how it was broken.
Nobody told it which message to attack #
According to Weierud, Carter Leffen contacted him on 15 September asking him to verify the result. Leffen’s only instruction to the model had been to see whether it could break any of the unsolved messages on the Crypto Cellar page. Astra looked through the list, settled on Nr. 172 as the most promising candidate, and suspected early that its plaintext was related to Nr. 173, SIPVX, a message from the same day that Alex Shovkoplyas broke back in 2017.
That hunch gave it a crib. Astra bet on the repeated place name “ROSENOW ROSENOW” appearing in the text, then wrote its own Enigma simulator and a software version of the Bombe in Python and C++, and ran the search. It worked. The whole thing took about two days. Weierud estimates a human researcher would have needed weeks, if not months, to do the same.
Why it stayed unsolved for so long #
This was not a matter of nobody having enough compute. A three-rotor Army Enigma is well within reach of a modern PC once you have a decent crib. The problems were elsewhere:
- The key was an outlier. MVUEH used wheel order 253, while the daily key and the SIPVX key both used 512. Anyone starting from the known keys for 10 July was looking in the wrong place.
- The transcription was dirty. The break showed that the ciphertext copied from the original message form contained several errors.
- The left-hand rotor turned over. It stepped at the 72nd letter, which rarely happens and is known to complicate a break.
- The operator made a typo. “Bitte” was enciphered as “Btte”, which, along with SIPVX repeating the sender’s signature, explains why the two nearly identical messages differ by twelve letters.
The part that surprised the archivist #
Weierud says he is still going through Astra’s logs, and one entry caught his attention. The model had traced the message corpus back to specific German Bundesarchiv radio-message volumes, RS 3-3/20a and RS 3-3/63b. Those references are correct, but they are not listed on the Crypto Cellar site. It noted that the original image of message Nr. 172 was still unlocated, and added, almost as an aside, that it had not sent any correspondence. Weierud himself spent weeks researching the same archive files.
It is also the second historical cipher Astra has cleared in quick succession: Tom’s Hardware notes it broke a 108-year-old WWI ADFGVX message last week. The raw cryptanalysis here is textbook, the same crib-and-bombe method Bletchley Park used. The new part is the model choosing the target, spotting the related message, writing its own tooling and chasing primary sources without being asked.
Weierud’s verdict is short: “as an old cryptanalyst, I am still in awe.”