A coded military letter linked to Napoleon had remained unread for 217 years. Carter Church reports that GPT-6 Astra turned a single scanned plate into a working French plaintext in about six hours of model execution. The result is striking, but the phrase “AI cracked the cipher” compresses several different jobs into one sentence.[1][2]
The system first had to segment 24 handwritten rows and transcribe roughly 1,300 cipher units made from numbers, letters and invented signs. It then connected the document to a previously published partial key, built a computational search over the remaining symbol assignments and checked candidate readings against French language patterns and known correspondence from March 1809.[1]
The workflow that produced the reading
| Stage | What happened | How it can be checked |
|---|---|---|
| Image transcription | The scan was divided into rows and handwritten signs were assigned consistent labels. | Compare the transcription directly with the surviving plate.Source 1 |
| Partial-key recovery | An earlier table supplied values for part of the cipher. | Verify each known assignment against the encoded text.Source 1 |
| Key search | A solver tested alternative symbol-to-letter assignments using simulated annealing and French-language scoring. | Rerun the published script and inspect whether the plaintext regenerates.Source 1 |
| Historical validation | The reading was compared with dated orders and campaign context, correcting the document from 1807 to 1809. | Check names, forces, dates and overlapping correspondence in independent archives.Source 1Source 2 |
Why the “single prompt” description is incomplete
The model reportedly began from one image and a goal, but the successful process still depended on external historical sources, an existing partial key and code generated to search the key space. That does not diminish the achievement. It identifies the real capability: coordinating perception, research, programming and iterative interpretation across a long task.[1]
A readable plaintext is only the start
Historical ciphers are unusually vulnerable to plausible nonsense. A language model can prefer a fluent French sentence even when several signs were mistranscribed. Church’s package leaves some signs uncertain and provides a script that regenerates the reading. That transparency lets cryptographers challenge individual assignments instead of accepting a polished translation on authority.[1]
Why archives may change faster than scholarship
Millions of pages contain hard handwriting, damaged scans, obsolete abbreviations and private ciphers. A model that can transcribe, write small tools and search contextual clues can expand the number of documents that receive a first pass. Human experts may then face a queue of proposed readings rather than a queue of untouched images.
That shift creates a new risk: an attractive solution can spread faster than curators can verify it. Archives will need versioned transcriptions, uncertainty marks, source images, reproducible scripts and review status. The model should accelerate the argument, not erase the evidence from which the argument was made.
The Marmont result is therefore best understood as an inspectable research workflow. AI generated and coordinated much of the search. Human historical knowledge, prior cryptographic work and independent checking turned the output into a proposed solution worth taking seriously.[1][2]
Sources & further reading
- Breaking the Marmont Cipher, 1809 Carter Church · 18 Sep 2026 · Accessed 16 Oct 2026
- AI helps crack historical ciphers and manuscripts Nature reporting republished by Folha de S.Paulo · 16 Oct 2026 · Accessed 16 Oct 2026