AI decodes Napoleon's cipher after 217 years in 6 hours

Two hundred and seventeen years. That's how long a coded military briefing to one of Napoleon's generals went without a full reading. An AI model reportedly got through it in about six hours.

Carter Church, an engineer at cybersecurity firm SentinelOne, gave OpenAI's GPT-6 Astra one low-resolution scan, a blurry reproduction from a 1969 French Army journal. In the AI decoding of Napoleon's cypher, the model both transcribed the marks and broke the code. Those two skills normally belong to different specialists.

How did GPT-6 Astra decode Napoleon's cypher?

The message runs to 24 lines. Astra logged about 1,300 hand-drawn marks and sorted them into 155 unique signs. It was a homophonic cypher, meaning several symbols can stand in for the same common letter.

The model then linked the letter to a partial key, catalogued by historian Daniel Tant, that nobody had tied to this message before. That key covered 33 letter values. Astra's own solver filled in the rest. Five signs stay ambiguous because they appear only once.

On March 16, 1809, he instructed his stepson Eugène, Viceroy of Italy, to send an encoded message to General Auguste de Marmont, whose troops in Dalmatia were isolated from the French army. The message was to be delivered by an intelligence officer.

The decoded message is an update on the position and size of French and allied armies before Austria's invasion, and verification of nearby letters helped Church to establish the correct date for publishing the letter.

An incorrect key cannot generate coherent and historically accurate text, as pointed out by Church, and therefore his solution is verifiable from a technical perspective.

Church has made available his transcriptions of the text, the decipherment tables and scripts. The entry in the list of unsolved historical cyphers at Cryptiana maintained by Satoshi Tomokiyo is labelled as solved.