
AI system cracks 217-year-old Napoleonic military cipher in six hours
Cybersecurity engineer Carter Church used OpenAI's GPT-6 Astra to transcribe and decode a 1,300-character letter sent to General Auguste de Marmont during the 1809 campaign against Austria.
The unsolved 1809 dispatch
In late March 1809, as France prepared for war against the Austrian Empire, a confidential military dispatch left the headquarters of Napoleon Bonaparte's stepson, Eugène de Beauharnais, the Viceroy of Italy. The message was directed across the Adriatic Sea to General Auguste de Marmont, whose forces were stationed in Dalmatia, in modern Croatia. Because Austrian territory separated Marmont from other French armies, an outbreak of hostilities risked cutting his corps off entirely.
The document opened with a single sentence in plaintext French: "Vous avez dû recevoir, Monsieur le Général Marmont, mes lettres des 8, 14 et 20 courant" (You must have received, General Marmont, my letters of the 8th, 14th and 20th of this month). Below that line lay 24 lines consisting of two-digit numbers, stray letters, and hand-drawn symbols, totalling 1,300 coded units. The cipher key was lost over time, leaving the contents unread for 217 years. Researchers possessed only a reproduction of the cipher table published in 1969 in the journal Revue historique des Armées.
- Napoleon instructs Eugene to dispatch military directives to General Marmont
- Eugene de Beauharnais sends the ciphered letter to Marmont in Dalmatia
- Revue historique des Armées publishes a reproduction of the cipher table
- Carter Church publishes the full decryption completed by GPT-6 Astra
Multimodal decryption process
In September 2026, Carter Church, an artificial intelligence engineer at the cybersecurity firm SentinelOne, applied OpenAI's GPT-6 Astra model to the problem. Working directly from the 1,202 by 1,836 pixel digital image of the 1969 reproduction, the system executed the entire transcription and decryption workflow in six hours.
What makes this impressive is not actually breaking the code, but that Astra completed the entire multimodal workflow in six hours, starting from a single image and a single goal.
Before solving the cipher, the model had to generate its own digital transcription. It segmented the image into rows, assessed character boundaries, and grouped visual variations. Out of 175 initially recorded hand-drawn marks, the system identified duplicate variants and reconstructed a 155-symbol homophonic substitution system. In this cipher system, individual letters could be represented by multiple distinct symbols, and certain glyphs represented complete words. Prior to this work, cryptographers had identified 33 symbols, covering 435 of the 1,300 units. The model resolved the remaining 865 units using simulated annealing to match symbol sequences with historical French vocabulary and syntax.
Astra sliced the 1,202 by 1,836 board into rows, read the glyphs, decided which of the 175 hand-drawn glyphs were the same glyph drawn twice, and only then began to crack the cipher.
- Previously known
- 435 units
- Solved by GPT-6 Astra
- 865 units
Strategic contents and historical verification
The deciphered text reveals critical strategic directives issued immediately before the outbreak of the 1809 campaign. The message detailed the dispositions and movements of allied Bavarian, Polish, and French units, while noting Russian troop movements. It instructed Marmont to prepare his troops to march and urged him not to be intimidated by his isolated position or by minor enemy forces. Marmont received the original orders in time, moved his army out of Dalmatia, and reinforced Napoleon's main force for the victory at the Battle of Wagram.
Following the publication of the decryption, Satoshi Tomokiyo, creator of the cryptography database Cryptiana, reviewed and verified the solution. The verification confirmed the reading and corrected the document's historical catalog entry, adjusting its estimated origin date from 1807 to late March 1809 based on the officer titles and troop deployments described in the text.


