GPT-6 Astra Breaks a 1941 Enigma Message That Stumped Codebreakers for Two Decades

An 82-letter German Army Enigma message that had resisted every attempt to break it since at least 2005 has been solved by OpenAI’s GPT-6 Astra, and the model picked the target, found the crib and wrote its own codebreaking tools, according to Frode Weierud, who keeps the archive of unbroken messages at Crypto Cellar Research.

Carter Leffen contacted Weierud in mid-September asking him to validate a break of message MVUEH, sent on July 10, 1941, and logged as No. 172 by an SS-Totenkopf supply unit’s radio station. “After Carter Leffen sent the MVUEH break details, it was immediately clear he had found the correct key and plaintext,” Weierud wrote.

What the model did on its own

Leffen’s only instruction was to see whether GPT-6 Astra could break any of the unbroken messages on Weierud’s site. The model chose MVUEH as the most promising, suspected early that its plaintext was related to message No. 173, SIPVX, which Alex Shovkoplyas broke in 2017, and after trying several approaches settled on the repeated place name “ROSENOW ROSENOW” as its crib. It then wrote an Enigma simulator and a Bombe-style attack in Python and C++ and ran the search that recovered the key.

The key turned out to be completely different from the day’s other keys, with wheel order 253 instead of 512. The break also showed that the published transcription of the ciphertext had several errors and that the machine’s left-hand wheel turned over at the 72nd letter, a rare event that makes a break harder. Weierud thinks those factors explain why no one had solved it before.

The plaintext itself is ordinary. It asks for the route of march, gives the sender’s position as Rosenow and requests an immediate radio reply. It’s nearly identical to SIPVX; the 12-letter difference comes from “Bitte” being enciphered as “Btte” and the sender’s signature, Waschbusch, being repeated in the later message.

Checking the work

The open-source AI safety project SWARM verified the result independently with its own Enigma I simulator, using none of the solver’s code and starting from the ciphertext as published in Weierud’s archive. Its output matched letter for letter. SWARM’s write-up points out that the published key is just a handful of settings anyone can type into a simulator, which makes the result easy to check.

Weierud and his colleagues are still going through the model’s logs. One shows it had traced the messages to specific Bundesarchiv radio-message volumes, RS 3-3/20a and RS 3-3/63b, references Weierud confirmed are correct even though they aren’t on his site. “What it has achieved in two days would take a human researcher weeks or even months,” he wrote. “The AI break of the Enigma message MVUEH is simply amazing, and as an old cryptanalyst, I am still in awe.”