GPT-6 Astra Cracks a 217-Year-Old Encrypted Letter From Napoleon

October 4 (IT Home) — An AI researcher has used OpenAI’s GPT-6 Astra to crack an encrypted military letter written by Napoleon Bonaparte in 1809. Because the codebook had been lost, the letter had never been fully deciphered in the past 217 years.

The work was done by Carter Church, an AI engineer at SentinelOne. He said he used the GPT-6 Astra model, starting from a scanned image of the historical document, to carry out text recognition and cryptanalysis on the letter; the whole process took about six hours.

The letter was written on 16 March 1809, when Austria was preparing for war with France. Napoleon wrote to his stepson, Eugène de Beauharnais, the Viceroy of Italy, asking him to pass military information to the French general Auguste de Marmont through encrypted letters. The surviving document begins with one line of ordinary French, followed by 24 lines of encrypted text made up of numbers, letters and invented symbols.

Researchers had previously established the meaning of 33 of the symbols, but because the complete codebook was lost, the rest could not be cracked. IT Home notes that the letter contains about 1,300 cipher units and 155 different symbols, some representing single letters and others corresponding to entire French words.

Church explained that GPT-6 Astra first analysed the scanned image, dividing the document into regions and identifying the symbols within them, and then tried to crack the cipher’s structure. The model used methods such as simulated annealing, running large numbers of tests on possible letter and word combinations for different symbols, and comparing them against patterns of letters and words in historical French texts.

During the analysis, the AI referred to features of historical French texts — including works by Alexandre Dumas, Victor Hugo and Marmont himself — to assess how plausible different decipherments were. The model also discovered that some symbols did not stand for individual letters but for whole French words.

The deciphered content revealed Napoleon’s military dispositions on the eve of the 1809 war with Austria. The letter lists the strength of French and allied forces in various parts of Europe, including armies in Bavaria, Poland, Saxony and Italy, as well as the situation of Russian troops advancing towards Austria.

The decryption also filled in some information missing from earlier historical records. The research shows that the letter to Marmont contains more detail about Austrian troop deployments, Marmont’s own position and French military plans. Marmont was then stationed in Dalmatia, far from Napoleon’s main army, so Napoleon wanted the letter to give him a picture of the overall situation and prepare him for action.

Church said that this kind of cracking work could in theory be done by hand, but usually requires a great deal of time and expertise. The distinguishing feature of this AI system is that it can carry out historical document recognition, symbol classification, cipher reasoning and text reconstruction — tasks previously handled separately by experts in different fields — all at once, and in only six hours.

To verify that the model was not simply generating answers from existing historical material, Church re-ran the analysis with historical texts about Napoleon and Marmont removed as references. According to him, the model still produced the same interpretation. The result was subsequently recognised by Satoshi Tomokiyo, the maintainer of the historical cipher research database Cryptiana, which has now marked the Marmont letter as a cracked cipher.