AI helped break a 1941 Enigma message left unsolved for 85 years

An 82-letter German Army radio message sent in July 1941, listed as unbroken in a public archive of Enigma traffic, has been read with help from OpenAI’s GPT-6 Astra. The archive now records it as broken, credited on 14 September 2026.

The result is unusual among AI claims because it is checkable end to end. The key, the code and the search data are published, so the decryption can be rerun by anyone, and the archive maintainer reviewed it before the listing changed. That combination, a specific historical artifact plus a reproducible method, is rare in this category.

How the message was read

Carter Leffen, who works at Bloomberg in New York, ran the investigation using GPT-6 Astra together with a set of specialist agents. His case study records 14,829,646 distinct keys checked against the message header across an investigation he describes as taking two days.

The decisive clue was not generated. It came from a different message sent the same day and already decrypted, which contained the town name Rosenow twice in succession, giving a crib to test against a message only 82 letters long. The English translation of the result reads: “Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio.”

The case study is careful about how much of that was compute. It says the two days “describe the span of the investigation,” that search processes ran in parallel, and that “the logs do not provide a complete total for human or model reasoning hours.” Leffen wrote separately that he “spent 99x more effort on building the website to describe the challenge and solution than any of my efforts on breaking the code.”

What has been confirmed and what has not

The 1941 message list now carries the message as broken, with Leffen credited on 14 September 2026. The case study builds on the archival work of Frode Weierud, who maintains CryptoCellar.

That listing records an accepted credit. It is not the same as an independent review, and none has been reported. What makes the claim testable anyway is that the key, the code and the search data were published rather than described, against a corpus that predates the model by 85 years.

What the episode does not establish is how much of the work the model did. The crib came from human archival knowledge, the pipeline was assembled by a person, and the case study declines to total the reasoning hours. The unresolved question is how many of the remaining unbroken messages in the archive yield to the same method now that it is documented.