A Napoleonic Letter Unread for 217 Years Speaks Again, With an AI Model’s Help

Carter Church had never broken a cipher before. A staff engineer at the cybersecurity company SentinelOne, he took on a letter to one of Napoleon’s generals that for 217 years sat in plain sight and said almost nothing. After one line of ordinary French come twenty-four rows of numbers, stray letters and invented symbols, known only from a plate in a French army journal from 1969.

In a write-up published September 18, 2026, Church gives a full reading of it. He used OpenAI’s GPT-6 Astra, which worked for about six hours of model time from that one scanned image. Satoshi Tomokiyo, the Japanese patent engineer and hobby cryptographer who keeps the Cryptiana list of unsolved historical ciphers, made a preliminary check, laying the decoded French over the original plate, and now lists it as solved. The result is a month old, but it moved again this week, when Nature reported on October 16 and AFP on October 17 that it is part of a much larger wave.

A general on the wrong side of the sea

The letter came from the headquarters of Eugène de Beauharnais, Viceroy of Italy, to General Marmont, whose Army of Dalmatia sat across the Adriatic, cut off by Austrian territory. Church redates it from the 1807 on the old listing to late March 1809, about two weeks before Austria invaded Bavaria. It tells Marmont where his side stood: forty thousand Bavarians at Munich and Passau, thirty thousand Poles on the Vistula, eighty thousand French at Bayreuth. It tells him that when the order to march comes, he must not be intimidated by a few troops or a gathering of rabble.

That last phrase is the small treasure. Napoleon’s own order of 16 March 1809, in print since 1865, breaks off at the same spot with a row of dots. The cipher letter carries the sentence from the receiving end, with no gap. Church is careful about what that proves: the letter is a paraphrase, so it does not pin down Napoleon’s exact word, but it is the first direct evidence of what the dots were hiding.

What the machine actually did

The model cut the plate into rows, read the signs and decided which of 175 hand-drawn marks were the same sign written twice. It found and used a partial key that the French cryptology historian Daniel Tant had posted online, whose 33 letter values covered 435 of the letter’s 1,300 cipher units. It worked out the rest with simulated annealing, which keeps adjusting a key and scores each attempt against the letter patterns of real French, and checked each proposed value against every place that sign appears. To rule out the model simply recalling history it had read, Church reran the solver with Marmont’s memoirs and every Napoleonic text removed. It recovered the same reading. The key, the transcription and scripts that regenerate the result are free to download.

Church notes the letter could have been solved without AI, Nature reports; the model cut the time and effort. Until now, he writes, no one who could read it could justify the time.

A flood of solutions, and a caution

Nature also describes Carter Leffen, a product development coach at Bloomberg with no cryptography experience, whose model tested more than four billion Enigma settings in ten hours to read a German message sent on 10 July 1941. Cryptiana’s notes now record solutions sent in with help from ChatGPT and Anthropic’s Claude, and Tomokiyo writes on his page on codebreaking with AI that the year 2026 will be remembered as a turning point for codebreaking of historical ciphers.

The people checking all this cannot keep up: Tomokiyo notes that he is receiving many solutions and has not caught up in listing them. Elonka Dunin, who has written books on cryptography, told Nature that many incoming AI solutions are false positives from people who do not understand cryptanalysis. A reading only counts once someone who knows the field has checked it.

Church’s rebuilt key now holds 155 signs where only 33 values were published before. If another letter in this system ever turns up, whoever finds it will not be starting from zero.