AI Helped Solve Two 1941 Enigma Messages

Two German Army Enigma messages from 1941 have been decoded after sitting unsolved in a research collection for about two decades. These were not cases of pasting ciphertext into a chatbot and accepting whatever came back. One model selected its own target and wrote tools to test a clue. The other worked inside a cryptanalyst’s prepared toolchain. Both results needed a human check.

An 82-letter message with the wrong-looking key

Developer Carter Leffen asked OpenAI’s GPT-6 Astra to investigate unbroken messages in the Crypto Cellar archive. It chose MVUEH, an 82-letter message dated July 10, 1941. A different message from the same day, SIPVX, had already been solved and contained almost the same text. That suggested a promising route, but the two messages did not share a key. MVUEH’s wheel order was 2-5-3; the other messages used 5-1-2.

Astra focused on the repeated place name ROSENOW as a crib—a likely fragment of the original text that could narrow the key search. It then built Enigma simulation and Bombe-style search software in Python and C++. Frode Weierud, who maintains the archive, verified the recovered key and plaintext.

There were extra obstacles. The available ciphertext transcription contained errors, and the Enigma’s left-hand wheel turned over at the 72nd character. Those details help explain why earlier attempts had failed; they also show why a clean-looking block of letters is not always clean data.

The second break had a stronger human hand

Cryptanalyst Jack Willis used Anthropic’s Claude Opus 5 on FMNGI, a message dated July 31, 1941. This time, the human supplied a Go-based cryptanalysis workbench and relevant archival material. A likely officer signature gave the search a 14-letter crib: XHARTJENSTEINX.

According to Weierud’s account of the FMNGI break, the final key search took 13 minutes and 28 seconds on an Apple M2 computer. The surviving received copy and archival transcription had errors of their own, so finding a key was only part of the job. The decoded text still had to make sense against the historical record. Weierud says seven Enigma messages in that collection remain unbroken.

Impressive research, not a warning about modern encryption

I wouldn’t treat these results as evidence that AI can suddenly read today’s encrypted chats. Enigma was a mechanical wartime cipher, and these attacks relied on related messages, likely words, archival records, custom search code, and specialist validation. The useful technical lesson is narrower: an AI system can speed up the tedious loop of finding clues, writing a simulator, testing candidates, and checking the result.

There is also a transparency limit. Weierud is still examining the Astra logs and says it is not clear how the model found references to some archival files. The plaintext may be verified; the full path the model took to get there is still being reconstructed.

Share: