An artificial intelligence model has successfully decoded a 217-year-old encrypted letter sent by Napoleon Bonaparte’s stepson to a French general in March 1809. Carter Church, an engineer at SentinelOne, used OpenAI’s GPT-6 Astra tool to transcribe and translate the historical document in roughly six hours from a single image and goal.
How GPT-6 Astra Cracked the Napoleon Cipher in Six Hours
Carter Church, an engineer at cybersecurity firm SentinelOne, fed a low-resolution scan of the manuscript into OpenAI’s new model, GPT-6 Astra, with a single prompt. The AI handled the entire multi-modal workflow end to end after Church found the image via a French army journal published in 1969.
During its independent research, the model identified a partial table by cryptology historian Daniel Tant that documented 33 of the symbols, accounting for roughly a third of the text. Although Tant had previously cataloged the table under a slightly different name, Astra made this connection itself via web research as it investigated, applying the key to decode the rest of the homophonic substitution system.
Decrypted Messages Reveal Napoleon's Troop Movements Before 1809 Defeat
The translated text outlines the precise positioning of French, Bavarian, and Polish forces on the continent just two weeks before Austrian troops launched their offensive. The message instructed Marmont to remain steadfast despite his isolation and assured him that enemy forces were heading toward defeat.
Using the decrypted troop counts, officer titles, and unit placements, Church re-dated the correspondence to late March 1809. Just two months later, on May 21, 1809, during the Battle of Aspern-Essling near Vienna, Austria, Napoleon suffered his first personal defeat, resulting in the death of 23,000 French soldiers.
Historical Value in Automated Cryptanalysis
Zack White, a history research fellow at the University of Portsmouth, called it a refreshing use of artificial intelligence to provide a credible solution to a historical mystery, noting that using identical tools and methodologies on unknown content increases trust in outcomes.
Church argued that while human cryptographers could have eventually solved the cipher given sufficient time, the immense multi-disciplinary labor required to transcribe, catalog, translate, and break historical codes leaves countless specialized problems untouched.