SentinelOne AI researcher uses GPT-6 Astra to crack Napoleon cipher after 217 years
Carter Church used AI to transcribe and decipher a coded 1809 letter to Napoleon’s general Marmont from a single scanned image.
An AI researcher at Israeli-American cybersecurity company SentinelOne has cracked a coded letter sent to one of Napoleon Bonaparte’s senior generals, deciphering a message that had remained unread for 217 years.
The breakthrough was achieved by Carter Church, a staff AI engineer at SentinelOne, who used OpenAI’s GPT-6 Astra to transcribe and decipher the letter to General Auguste de Marmont. Church said the entire process took about six hours, starting from a single scanned image of the document.
The letter was written in March 1809, as Austria prepared to go to war with Napoleon’s France. Napoleon had written to his stepson, Eugène de Beauharnais, on March 16 and specifically instructed him to send Marmont a letter in cipher, carried by an “intelligent officer.” The surviving document contains one line of ordinary French followed by 24 rows of numbers, letters and invented symbols.
The cipher’s key was lost, leaving the letter on lists of unsolved historical codes for decades.
The message contains about 1,300 cipher units and 155 distinct signs. Researchers had previously identified the meaning of 33 of the signs, accounting for roughly a third of the encrypted material, but the rest had remained undeciphered.
Church said GPT-6 Astra first analyzed the scanned image, cutting it into sections and identifying the symbols before attempting to break the cipher. It then used a technique known as simulated annealing, essentially a large-scale trial-and-error search, to test possible assignments of letters and words to the remaining symbols.
The model compared candidate French texts using letter and word patterns drawn from historical French writing, including works by Alexandre Dumas, Victor Hugo and Marmont himself. It also identified signs that represented entire French words rather than individual letters.
The resulting solution produced a complete reading of the letter.
The breakthrough was inspired by another recent example of AI being used to decipher a long-unsolved military cipher. Church said he had seen a report about an AI model deciphering a coded German military message and decided to see whether the same approach could be applied to the Marmont letter.
The deciphered document sheds light on Napoleon’s military planning in the weeks before Austria invaded Bavaria in April 1809.
Napoleon’s original instruction to Eugène had been to tell Marmont where French and allied armies were positioned. The deciphered letter lists French and allied forces across Europe, including troops in Bavaria, Poland, Saxony and Italy, as well as Russian forces marching against Austria.
It also reveals information that was missing from the surviving printed version of Napoleon’s March 16 instructions. Church’s reconstruction indicates that the letter contained additional information about Austrian forces and Marmont’s own position, as well as a fuller version of a sentence that appears incomplete in the historical record.
The message was intended for Marmont, who was stationed with French forces in Dalmatia, on the other side of the Adriatic from Napoleon’s main armies. Napoleon was preparing for a conflict with Austria and wanted Marmont to understand the broader military situation and be ready to act.
The letter’s importance therefore goes beyond the novelty of its encryption. It provides a glimpse into how Napoleon’s headquarters communicated with an isolated commander immediately before the outbreak of one of the major wars of the Napoleonic era.
Church said the process could theoretically have been performed manually, but would have required extensive and painstaking work. The significance of the AI breakthrough was that the same system was able to handle tasks that would traditionally have been separated between document transcription and cryptanalysis.
The model first had to determine which marks in the scanned document represented the same symbols. It then had to infer the meaning of unknown symbols, test those hypotheses against the rest of the document and repeatedly refine the resulting French text.
Church also reran the analysis while removing Napoleonic and Marmont-related material from the model’s available historical texts. According to his account, the system recovered the same reading, an important check against the possibility that it was simply reproducing information it had encountered elsewhere.
The solution was subsequently recognized by Satoshi Tomokiyo, whose Cryptiana database had listed the Marmont letter among unsolved historical ciphers. Tomokiyo has now marked the cipher as solved.
The episode also highlights the growing capabilities of AI systems in areas traditionally associated with specialist human expertise. GPT-6 Astra is designed for advanced reasoning, computer use and professional tasks, and OpenAI says the model can handle complex multi-step workflows.
But the same capabilities that allow AI to recover a 217-year-old military message raise questions for cybersecurity. Systems capable of recognizing patterns, reconstructing missing information and automating sophisticated analysis could also potentially be applied to modern encrypted or obfuscated data.
That dual-use potential is particularly relevant for researchers such as Church, whose work sits at the intersection of artificial intelligence and cybersecurity.
In this case, however, the result was not a new military secret. It was a message that had been sitting in plain sight for more than two centuries, waiting for someone to reconstruct the key.
