Technology · Article

How AI Cracked a 217-Year-Old Napoleon Military Cipher in Six Hours

Published:

The Emperor Napoleon in His Study at the Tuileries label QS:Len,"The Emperor Napoleon in His Study at the Tuileries" label QS:Lar,"الإمبراطور ناپُليُون في مكتبه بِقصر التويليري" title QS:P1476,en:" The Emperor Napoleon in His Study at the Tuileries label QS:Len,"The Emperor Napol
Image credit: Jacques-Louis David · Public domain · Wikimedia Commons

In brief

An AI engineer at the cybersecurity firm SentinelOne used OpenAI's GPT-6 Astra model to decode a homophonic-substitution cipher that Napoleon Bonaparte's headquarters sent to General Auguste de Marmont in March 1809. Working from a single scanned image, the system transcribed, analyzed and solved the 217-year-old code in roughly six hours, a task that had defeated cryptographers for more than two centuries.

A military cipher sent on Napoleon Bonaparte's orders in 1809 has been fully read for what researchers say is the first time since it was written. Carter Church, an AI engineer at the cybersecurity company SentinelOne, used OpenAI's GPT-6 Astra model to transcribe and decode the message, according to Türkiye Today and Euronews. The Star identified Church as 27 years old. The breakthrough took roughly six hours of computer time, and it is part of a wider wave of AI-assisted solutions to historical ciphers reported in October 2026.

A homophonic cipher that defeated historians for two centuries

The document is a homophonic substitution cipher, a system in which several different symbols can stand for the same letter, which makes brute-force analysis far harder than with a simple substitution code, Türkiye Today reported. According to the technology outlet Cryptonomist, the encrypted text runs across 24 rows and contains roughly 1,300 cipher units drawn from about 155 distinct symbols. Before this project, researchers had identified the meaning of only 33 of those symbols, Cryptonomist reported, leaving most of the message unread since it was written.

The letter had sat on a public list of unsolved historical ciphers compiled by Satoshi Tomokiyo, a Tokyo-based researcher who runs the Cryptiana cipher-history website, according to a report carried by The Star of Malaysia. Tomokiyo verified Church's solution, the report said.

How GPT-6 Astra solved it in about six hours

Church fed the AI model a single scanned image of the document along with one prompt, and GPT-6 Astra then carried out the entire workflow itself, Cryptonomist reported. The model transcribed the handwritten symbols into machine-readable text, analyzed repeating patterns in the substitution system, wrote its own code to test candidate solutions, and used a trial-and-error search method known as simulated annealing to try multiple word fits at once. It cross-checked candidate decodings against period French texts, including writings associated with Marmont himself, before settling on a final reading, according to the same report.

To rule out the possibility that the model was simply recalling memorized historical text rather than genuinely decoding the cipher, Church reran the analysis after removing Napoleon-related reference material from the process, Cryptonomist reported. He published his full research record, including the transcription, cipher tables and scripts, so that other researchers could reproduce the result, according to Euronews.

What the decoded letter says

The message originated from the headquarters of Eugène de Beauharnais, Viceroy of Italy and Napoleon's stepson, and was addressed to French General Auguste de Marmont, according to both Euronews and Türkiye Today. Türkiye Today reported that Napoleon instructed the dispatch to be sent on 16 March 1809. Euronews reported that Church used decoded details such as troop numbers and officers' titles to redate the document to late March 1809, about two weeks before Austrian forces launched their offensive.

The decoded text sets out troop positions, listing Bavarian, Polish and French forces and noting 80,000 troops in the Friuli army, Türkiye Today reported. It also indicates that Russian forces were moving against Austria and encourages Marmont to advance without hesitation once ordered, rather than being deterred by smaller opposing forces, according to Türkiye Today. Euronews reported that, at the time, large French forces were tied down fighting in Spain, leaving Marmont's corps in Dalmatia at risk of isolation as Austria prepared to act. The Austrian campaign that followed eventually ended in a French victory at the Battle of Wagram, according to a report from Jfeed.

The cipher before and after AI analysis
MetricBefore Church's projectAfter GPT-6 Astra
Symbols identified33 of roughly 155 (Cryptonomist)All roughly 155 symbols (Cryptonomist)
Time to solveUnsolved for 217 yearsAbout 6 hours (Türkiye Today, Euronews)
MethodManual cryptanalysis by historiansSingle image, single prompt, AI-run transcription, pattern analysis and simulated annealing (Cryptonomist)

Part of a wider pattern: AI and historical ciphers

Church's result is not an isolated case. Since mid-September 2026, most of the remaining entries on Tomokiyo's Cryptiana list of unsolved historical ciphers have been cracked with AI assistance, according to the report carried by The Star. Tomokiyo described 2026 as what will be remembered as a turning point for the codebreaking of historical ciphers, the same report said.

I'm excited long-time problems are now solved, but on the other hand, I have mixed feelings.

— Satoshi Tomokiyo, Cryptiana cipher-list curator, quoted by The Star

The same report said an unnamed cryptographer told Tomokiyo that the ease with which AI had solved these problems made him feel his own skills were no longer needed. The reaction reflects a tension in the field: AI tools are letting people with little specialist training in cryptography or 19th-century history solve puzzles that previously required years of dedicated study by historians and hobbyists, according to The Star.

Church himself said the most significant part of the project was not the codebreaking itself but the fact that GPT-6 Astra completed the full multi-step workflow — transcription, cryptanalysis, coding and historical cross-referencing — from a single image and a single goal, in about six hours, as reported by Cryptonomist.

Questions

Who decoded Napoleon's 217-year-old cipher?
Carter Church, a 27-year-old AI engineer at the cybersecurity firm SentinelOne, decoded it using OpenAI's GPT-6 Astra model, according to Türkiye Today and The Star.
What kind of cipher was used in the letter?
It is a homophonic substitution cipher, in which multiple symbols can represent the same letter, making it harder to solve than a simple substitution code, Türkiye Today reported.
How long did the AI take to crack the code?
GPT-6 Astra completed the transcription, analysis and decoding in roughly six hours after being given a single scanned image and one prompt, according to Euronews and Cryptonomist.
What does the decoded letter say?
It reports troop positions of Bavarian, Polish and French forces, including 80,000 troops in the Friuli army, notes Russian forces moving against Austria, and urges General Marmont to advance without hesitation once ordered, according to Türkiye Today.
Is this the only historical cipher AI has solved recently?
No. According to The Star, most of the remaining entries on researcher Satoshi Tomokiyo's list of unsolved historical ciphers have been cracked with AI assistance since mid-September 2026.

Sources

  1. Türkiye Today — AI cracks Napoleon's 217-year-old secret military letter in just 6 hours
  2. Euronews — AI decodes 217-year-old Napoleonic letter in 6 hours
  3. Cryptonomist — GPT-6 codebreaking AI cracks 217-year-old Napoleon cipher in six hours
  4. The Star — 'Mixed feelings': AI cracks centuries-old cyphers
  5. Jfeed — Israeli AI Cracks Napoleon's 217-Year-Old Secret Code

This piece was prepared with artificial intelligence and passed an automated source check before publication. If you see an error, ask for a correction. Request a correction or reply · How we make this

Related