For eight decades, a cryptic 82-character radio transmission from July 1941 sat untouched in historical archives. It was just another unsolved Wehrmacht message, tagged MVUEH, lost among millions of intercepted signals. Now, Carter Leffen, a product development coach at Bloomberg LP, has finally broken the silence. He didn’t use a room full of mathematicians or custom-built supercomputers. Instead, he leaned on OpenAI’s GPT-6 Astra, specifically the “Extra High” variant, to do the heavy lifting in roughly ten hours.
The process wasn’t magic; it was methodical. The Enigma machine offered 159 quintillion possible daily settings, making brute force impossible. Leffen’s AI agents scoured historical records and found a crucial clue: a previously decrypted message from the same day mentioned the town “Rosenow” twice. Guessing this location appeared in the unsolved text provided a foothold. The AI then exploited a known Enigma flaw—the machine never encrypts a letter as itself—to eliminate incorrect positions. Once the settings aligned, the remaining characters formed coherent German, including requests for marching orders and immediate radio replies.
What makes this breakthrough compelling isn’t just the solution, but its imperfections. The decrypted text contains typos like “BTTE” instead of “BITTE.” Leffen argues these errors prove authenticity, as a fabricated result would likely be polished. He notes that building the website to document the find took “99 times more effort” than the actual cryptanalysis. The AI handled everything from writing simulation code to cross-checking keys, while Leffen directed the strategy.
This case highlights a shift in how we approach historical puzzles. AI isn’t replacing human intuition here; it’s amplifying it. Leffen built the framework, but the model executed the tedious trial-and-error at scale. While independent experts must still verify the findings against the published 3D simulator and code, the speed is undeniable. If a single developer can undo 83 years of secrecy in a workday, we need to rethink what other “unsolvable” historical data might be waiting for the right prompt.
