GPT-6 Astra helped decode a 217-year-old cipher letter to Napoleon's marshal

Why it matters

The case shows how multimodal AI can combine image transcription, historical research and code-based analysis on a neglected archival problem, while the released key and script make the proposed result open to checking.

On September 18th, Carter Church published a proposed solution to a letter addressed to Napoleon's marshal Auguste de Marmont, saying GPT-6 Astra helped transcribe and decode the cipher in about six hours of model execution. Church's full write-up includes the key, transcription and a script intended to reproduce the reading. He recapped the work on X on September 30th.

The letter's 24 rows of symbols had appeared in a 1969 French military-history journal article, but the ciphertext was not transcribed in the historical-cipher listing that had cataloged it as unsolved. Church used the scanned plate in J. Vilcoq's article, rather than a ready-made text. He says Astra read the image, reconciled handwritten variants, incorporated an existing partial key and used a simulated-annealing solver to assign values to the remaining signs.

Church's account counts 1,300 cipher units and 155 distinct signs. A partial table compiled by French cryptology historian Daniel Tant supplied 33 letter values covering about 435 units, according to Church. The rest of the proposed solution depends on interpreting the scanned letter. Church says the model ran the work from the single image and a goal, while also noting that the cipher itself was not especially difficult for a specialist with time to work through it.

The recovered French text is a military briefing from Eugène de Beauharnais's headquarters, reporting troop positions and Austrian movements shortly before Austria invaded in April 1809. Napoleon had instructed Eugène on March 16th to send Marmont, then commanding French forces in Dalmatia, an encrypted account of the armies. The deciphered letter gives Marmont that briefing and tells him not to be intimidated by "a few troops or a gathering of rabble." It also fills out a sentence that breaks off in Napoleon's 1865 memoir, according to Church's comparison of the two texts.

Church says Satoshi Tomokiyo, who maintains the historical-cipher site Cryptiana, reviewed the proposed solution and marked the entry solved with its date corrected to 1809. That is meaningful corroboration from the site's maintainer, while the published files make the proposed key and output available for scrutiny. Church also identifies five signs whose readings remain uncertain, including alternatives for a reference to "His Majesty" and the word translated as "rabble"; he says those ambiguities do not change the letter's overall meaning.

https://x.com/CarterWChurch/status/2105167567706816690

The distinction between recovering a plaintext and establishing every historical detail matters here. Church's published decoder can test whether the proposed key maps the transcribed signs to the claimed French text. His historical dating also draws on the letter's troop information and comparison with Napoleon's correspondence. The result is a worked example of AI-assisted archival research, not evidence that a model independently authenticated the document or that every interpretation is settled.

Church brings relevant experience in building agents for a very different setting. His professional biography says he spent more than a decade in security operations before moving into AI engineering at SentinelOne, where he built NEON, an agentic framework for investigations and analyst support. That background helps explain the emphasis on an end-to-end workflow and a reproducible result: the cipher-breaking step was only one part of a process that also required image reading, historical research, translation and checking.

The GPT-6 Astra announcement describes the model's computer-use and multimodal capabilities, but the Marmont result remains Church's project and reported experience. He has released the materials needed to examine his method; the claims about the model's roughly six hours of execution and the historical interpretation are his account. The notable feat is that one workflow connected a hard-to-read archival image to a proposed, testable reading of a document that had received little attention.