OpenAI model decodes 217-year-old Napoleonic cipher from blurry scan

Cracking a Napoleonic Secret

An OpenAI model has successfully decoded a 217-year-old cipher sent by Napoleon Bonaparte. The document was a letter directed to General Auguste de Marmont, whose forces were trapped in Dalmatia in March 1809. Napoleon wanted the message encrypted so it would remain secret if captured by enemies.

The source material was a single low-resolution scan from a 1969 French Army journal. Despite the blurry quality, the model was able to reconstruct 24 lines of symbols, numbers, and letters. This effort revealed a war briefing detailing the positions of French and allied armies before the Austrian invasion.

The achievement was significant enough that the editor of Cryptiana's Unsolved Historical Ciphers list has marked the entry as solved. Carter Church, an AI engineer at SentinelOne, spent about six hours verifying the results. He noted that while human cryptographers could have solved it with enough time, the AI provided a unique automated solution.

How the AI Processed the Code

The model used to solve the puzzle was GPT-6 Astra. It began by logging approximately 1,300 hand-drawn marks from the image and sorting them into 155 unique signs. This was a necessary first step before the system could begin the actual decoding process.

The letter utilized a homophonic cipher, which is a method where different symbols can stand in for the same common letter. To handle this complexity, the model used a simulated annealing solver to complete the decipherment. It also linked the letter to a partial key that had not been previously associated with it.

One of the most notable aspects of the process was the unified workflow. Usually, transcribing text and performing cryptanalysis are separate skills or careers, but the AI performed them as one continuous task. The system checked proposed values back against every occurrence of a sign on the image to ensure consistency.

Correcting the Historical Record

Beyond just decoding the secret message, the model also helped fix a historical error. By checking nearby letters, the AI was able to correct the published date of the original letter. This shows the model's ability to use surrounding context to validate specific data points.

The decoded text confirmed that the dispatch was sent out before Austria invaded. It provided a clear picture of the military situation in Dalmatia at the time. This level of detail allows historians to see exactly what Napoleon was communicating to his stepson, the Viceroy of Italy.

Church emphasized that the solution was mechanically testable. Because the output was coherent and historically consistent, there was essentially no chance that an incorrect key could have produced the result. This provides a high level of confidence in the AI's ability to handle niche, unsolved problems.

The Future of Data Retrieval

This success highlights a shift in how we might approach historical research. Church noted that many fields have a backlog of niche problems that specialists simply do not have the bandwidth to tackle. AI systems can now act as a tool for outsiders who are simply curious to explore these archives.

The technology used to process an 1809 manuscript has broader implications for modern industries. For example, similar systems could help utilities forecast energy demand or optimize clean energy on the power grid. These applications rely on the same ability to process complex, multi-variable data sets.

However, these advancements come with significant environmental and social considerations. The computing boom required to run these powerful tools can consume vast amounts of electricity and water. There are also ongoing concerns regarding cybersecurity, potential misuse, and rising energy costs for households due to data-center demand.

The Human Element in AI

While the AI did the heavy lifting of decoding the cipher, human oversight remained essential. Church spent hours verifying the model's output to ensure it met historical standards. This partnership between human expertise and machine processing is becoming a standard for complex data analysis.

The project serves as a case study for how generative models are moving beyond simple chat interactions. They are now being used as sophisticated engines for