AI researcher Carter Church has deciphered an encrypted military briefing sent to Auguste de Marmont, one of Napoleon Bonaparte’s generals, using OpenAI’s deep reasoning model GPT-6 Astra. The document had remained undeciphered for 217 years. Starting from a single-page, low-resolution scan published in a French military journal in 1969, the model autonomously completed the steps of image processing, symbol mapping, cryptographic solver coding, and historical cross-validation in about six hours.
What made the letter undecipherable for two centuries?
The letter appeared in an article published by military historian Lieutenant Colonel J. Vilcoq in Revue historique des Armées in 1969. The first line of the 1,202 x 1,836-pixel black-and-white scan was plain French text; the following 24 lines contained about 1,300 cipher units made up of numbers, Greek letters, and hand-drawn symbols.
Two factors made the cipher unbreakable. First was the homophonic substitution method, in which the French headquarters clerk assigned multiple signs to frequently used letters; for example, nine separate symbols were used for the letter ‘E’ and eleven for ‘A’. That is why classic letter-frequency counting did not work. Second was the loss of the word table: the partial table prepared by researcher Daniel Tant gave equivalents for only 33 letters, while all the whole-word symbols in the text remained blank. In the Cryptiana archive, which cryptanalysts use as a reference, the letter was listed in the category of undeciphered ciphers.
A six-hour autonomous workflow
Astra carried out several stages in sequence without human intervention. First, by slicing the archive image line by line, it separated ink blots and similar glyphs in the handwriting; it eliminated 175 temporary labels and established that the document contained 155 unique cipher symbols. Then, taking the known 33 letters and the phrase ‘CONSEQUENT’, left in plaintext on line 14, as its core, it laid the groundwork for about one third of the text.
To guess the remaining symbols, it wrote a custom cryptanalysis solver in Python based on simulated annealing; it scored candidate letter combinations using French n-gram statistics compiled from the memoirs of Alexandre Dumas, Victor Hugo, and Marmont. It turned out that some symbols that constantly disrupted grammatical rules represented not single letters but 29 whole French words such as ‘de’, ‘que’, ‘les’, ‘vous’, and ‘général’. Finally, the model retrieved Marmont’s memoirs and Napoleonic-era texts from its database and reproduced the same solution; in this way it confirmed that the result came from cryptanalytic computation, not memorization.
What did the deciphered text add to history?
The biggest surprise came in the dating. The document, believed for two centuries to date from 1807, was actually written at the end of March 1809, two weeks before Austria invaded Bavaria and started the War of the Fifth Coalition. The letter had been sent by Napoleon’s stepson and Viceroy of Italy, Eugène de Beauharnais, to Marmont, who was isolated behind enemy lines in Dalmatia with 13,000 troops; the text aimed to boost the general’s morale by listing the positions of other army units one by one.
The solution also closed a 160-year-old debate. A sentence in Napoleon’s order dated March 16, 1809, in the official correspondence compilation published in 1865 had been cut, replaced by a row of dots. In the text Astra deciphered, the original of that sentence emerged: Napoleon was telling Marmont that he must not let a few soldiers—or ‘a handful of rabble’—intimidate him. Satoshi Tomokiyo, who manages the Cryptiana archive, also updated the document’s status to ‘Deciphered’ and corrected its date to 1809.
According to Carter Church, the real significance of the work lies not in cryptography but in AI’s impact on research methodology. Church says that none of the material in the letter was lost; that examining 1,300 symbols one by one, compiling 19th-century French statistics, and building a solver would take an expert weeks. GPT-6 Astra reduced this discovery cost to six hours. The fact that an undeciphered 1941 Nazi Enigma message was also solved with the same model in recent weeks shows that similar work in archival science will accelerate.
Source: WebTuring

Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.