On July 10, 1941, a German Army radio operator sent an 82 letter message encrypted with Enigma. The message, identified as MVUEH, appeared in a public collection of unsolved ciphers in 2005. In September 2026, researcher Carter Leffen used GPT 6 Astra to investigate it. Cryptanalyst Frode Weierud subsequently confirmed that the recovered key and plaintext were correct.
Leffen gave the model a broad goal: examine the unsolved messages and try to break one. Astra selected MVUEH, noticed that another message from the same day might contain related text, and used the repeated place name “ROSENOW ROSENOW” as a phrase to test. It developed software to simulate Enigma and search for possible settings.
Several details had complicated earlier attempts. The recovered rotor order differed from the known settings for other messages sent that day. The published ciphertext also contained transcription errors, and the left rotor moved at the message’s 72nd letter. The resulting plaintext asks for a route of march, reports the sender’s location as Rosenow and requests an immediate radio reply.
The investigation spanned September 14 and 15. The project describes it as work led by a researcher with AI agents, software searches and independent checks. It does not establish an exact number of hours spent reasoning by the model. This is a striking example of AI helping with historical research: connecting records, writing tools, testing a hypothesis and checking the answer against surviving evidence.
Explore AI models for your own projects at opencheese.dev