A military message sent on Napoleon Bonaparte’s orders in 1809 has reportedly been deciphered using artificial intelligence, revealing troop positions and strategic instructions that had remained encrypted for more than two centuries.

Carter Church, an engineer at cybersecurity company SentinelOne, reported that OpenAI’s GPT-6 Astra model completed the task in roughly six hours. According to his account, the model worked through the entire process: reading the scanned document, identifying handwritten symbols, analysing the cipher, writing code and checking the result against historical records.

The letter originated at the headquarters of Eugène de Beauharnais, Napoleon’s Viceroy of Italy, and was addressed to General Auguste de Marmont. Although written on Napoleon’s instructions, it was not a letter penned by Napoleon himself.

Turning handwritten symbols into readable text

Before attempting to decipher the message, the model had to establish what was actually on the page.

According to Church, Astra divided the scanned image into rows and examined 175 handwritten marks, determining which represented distinct symbols and which were variations of the same sign. Those distinctions mattered: an inaccurate transcription could undermine the subsequent analysis.

The message used a homophonic substitution cipher, in which multiple symbols can represent the same letter. This makes straightforward frequency analysis more difficult than with a basic substitution cipher.

Church reported that Astra combined image interpretation, cipher analysis and code execution to reconstruct the message. Historical records provided additional context for assessing whether the decoded names, troop numbers and locations made sense.

Researchers had examined the document over the decades, but, according to the account, no complete decipherment had previously been published.

A military briefing before Austria’s offensive

The decoded message describes French and allied troop deployments as Austria prepared for war with France in 1809.

With substantial French forces committed to Spain, Marmont’s troops in Dalmatia faced the risk of isolation. On March 16, Napoleon instructed Eugène to communicate the positions of the main French and allied forces to Marmont through an encrypted message carried by a trusted officer.

The recovered text outlines the positions of Bavarian, Polish and French forces. It states that the army in Friuli numbered 80,000 men, refers to Russian movements against Austria and urges Marmont to advance confidently when ordered.

These statements reveal the information and assurances being communicated to Marmont at the time. They should be read as the contents of a military briefing, rather than an independently verified account of every deployment.

Using the troop figures, officers’ titles and unit locations, Church dated the document to late March 1809, approximately two weeks before Austria launched its offensive.

Why the research record matters

Church published his research record on September 18, including the transcription, cipher tables and scripts intended to let others reproduce the result.

He also reported that Satoshi Tomokiyo, the researcher behind the historical cryptography website Cryptiana, reviewed the solution and marked the document as solved.

That supporting material is central to evaluating the claim. A convincing translation alone provides limited evidence. A documented path from the original symbols to the decoded text gives other researchers a way to inspect assumptions, identify errors and test the proposed solution.

The practical lesson for technical teams

For IT and business leaders, the case illustrates the potential of AI workflows that connect several capabilities within one task: interpreting an image, structuring uncertain data, generating code and checking outputs against external evidence.

Similar sequences appear in work involving scanned archives, legacy documentation and poorly structured operational records. Their usefulness depends on maintaining traceability between the source material, intermediate decisions and final output.

The reported result concerns a historical cipher; it does not establish that AI can break modern encryption. Its broader relevance lies in how a complex research problem was approached—and in the availability of evidence that allows others to assess the answer.

Source