A coded military letter from the Napoleonic era, which had sat for decades on lists of unsolved historical ciphers, can now be read in full. Engineer Carter Church, an AI specialist at cybersecurity company SentinelOne, used OpenAI's GPT-6 Astra model to decipher it in about six hours of model execution. The case, published by Church on his website and on X, has been picked up this week by Live Science, Euronews and Infobae, among other outlets.
The document was written at the headquarters of Eugène de Beauharnais, Napoleon's stepson and Viceroy of Italy, and addressed to General Auguste de Marmont, who commanded the French troops in Dalmatia on the far side of the Adriatic. Its first line is in plain French: Marmont should have received the letters of the 8th, 14th and 20th. Then come 24 rows of two-digit numbers, stray letters and invented symbols, whose key did not survive.
The historical context explains its importance. In February 1809, Austria decided to go to war against France, taking advantage of the fact that part of the French army was still tied down in Spain. Marmont, with about 13,000 men by Napoleon's own count, was isolated: Austrian territory lay between him and every other French army. On March 16, 1809, Napoleon ordered Eugène to inform Marmont of the position of every allied force “in a ciphered letter and by an intelligent officer.”
The only available image of the letter was a low-resolution plate published in 1969 in the Revue historique des Armées, digitized by the French academic archive Persée, measuring 1,202 by 1,836 pixels. According to Church, the text contains 1,300 cipher units and 155 distinct signs. It is a homophonic substitution cipher, in which the most common letters have several interchangeable signs so that simple frequency counting is useless.
The most striking point is that all the work started from a single image and a single prompt. No prior transcription existed, so Astra cut the plate into rows, read the signs and decided which of 175 hand-drawn marks were really the same sign written twice before it began breaking the cipher. Church stresses that transcription and cryptanalysis are usually separate skills, often separate careers, and that here they ran as one workflow.
The model also found a key piece on its own. As Church told Live Science, Astra used web searches to locate a partial table compiled by cryptology historian Daniel Tant and hosted by the French association ARCSI, which documented 33 letter values. The table was catalogued under a name that may contain a typo, “Grammont” instead of “Marmont.” Those values covered about 435 of the 1,300 units, roughly a third of the text.
The rest was obtained with a reproducible method. A program assigned letters to the remaining signs using simulated annealing, scoring each candidate text against French statistics drawn from works by Victor Hugo, Alexandre Dumas and Marmont's own memoirs. Twenty-nine signs turned out to stand for whole words, such as “de,” “que,” “vous” or “général.” To rule out bias, the process was rerun with every Napoleonic text removed from the language model, and it produced the same reading.
The deciphered content describes the disposition of French and allied forces weeks before the Austrian offensive: the Duke of Danzig with 40,000 Bavarians at Munich and Passau, Prince Poniatowski with 30,000 Poles on the Vistula, the Duke of Auerstaedt with 80,000 French troops at Bayreuth and the Duke of Rivoli with 60,000 men at Ulm and Donauwörth. One sentence states that “the Russians are marching against Austria, and the army of Friuli today numbers eighty thousand men.”
The letter ends with a direct warning to Marmont: “When you receive the order to march, you must not allow yourself to be intimidated by a few troops or a gathering of rabble.” Based on the troop figures, titles and positions, Church dated the document to late March 1809, about two weeks before Austria crossed the Inn River on April 10, and corrected the 1807 date under which it had been catalogued.
External validation has been key. Satoshi Tomokiyo, a Tokyo-based researcher who maintains Cryptiana, a repository of unsolved historical ciphers, reviewed the solution and marked the entry as “solved.” Church also published a package with the transcription, the 155-sign key and the scripts needed for anyone to reproduce the reading. “The solution is mechanically testable,” he told Live Science: an incorrect key would not produce coherent, historically consistent text.
The historians consulted by Live Science welcomed the result with nuance. Michael Rowe of King's College London called it “impressive” and compared it to “a kind of Rosetta Stone”: because Napoleon's original order survives, it is possible to check that the AI reproduces what was expected. But he also noted that “in a sense it doesn't tell us anything new,” since the cipher repeats much of that order. Zack White of the University of Portsmouth welcomed seeing AI used productively.
Church highlights one detail that does add something new. In the printed edition of Napoleon's correspondence, published in 1865, the sentence about not being intimidated ends in a row of dots. The cipher letter completes that idea from the receiving end with the phrase “un rassemblement de canailles,” although he himself warns that it is a paraphrase and not proof of the emperor's exact words. He also acknowledges five signs with alternative readings that do not change the meaning.
Beyond the historical anecdote, the case shows how models with vision, search and coding abilities can combine into a single workflow tasks that used to require several specialists. Church argues that the necessary material had been public for decades, but that no one could justify weeks of expert work on a single plate. According to him, Tomokiyo is now receiving solutions faster than he can record them, many from people who are not cryptographers.
In conclusion, deciphering the letter to Marmont is a useful example of AI applied to research because it comes with evidence: a documented method, reproducible files and review by an independent expert. That is the standard any similar claim should meet. As Church put it, generating possible solutions is easy; the hard part is verification, because “you simply cannot automate accountability.”
