AI
“🔐 Napolyon dönemine ait ve 217 yıldır çözülemeyen 1809 tarihli gizli askeri savaş brifingi, OpenAI'ın GPT-6 Astra yapay zeka modeliyle sadece 6 saat içinde deşifre edilerek tarihsel bir gizem aydınlatılmıştır. Bu başarı, yapay zekanın geçmişin kilitli arşivlerini açan güçlü bir dijital anahtar haline geldiğini ve tarih ile kriptoloji…”
Plain restatementAn encrypted military briefing dated 1809, from the Napoleonic period, that had been listed as unsolved for 217 years was deciphered in about six hours using OpenAI's GPT-6 Astra model. The post further asserts that this shows AI is now a general tool for opening historical archives and that a new era in historical and cryptological research has begun.
The core of this post is real. An encrypted 1809 military briefing, sent to Napoleon's General Marmont on the emperor's orders and long listed as unsolved, was read for the first time using OpenAI's GPT-6 Astra, and the researcher who did it published a full solution with the data and scripts needed to check it. The cipher catalogue that had listed the letter as unsolved is reported to have reviewed the result and marked it solved. Two things in the post are overstated. The six hours refers to model runtime only: the engineer behind it says he also spent several evenings on the project, and a third of the cipher's values had already been documented by earlier researchers. He also said plainly that the cipher itself was not hard and that human cryptographers could have broken it given the time, so the obstacle was mainly that the only available image was a single blurry 1969 reproduction filed under the wrong year. The closing claim that AI is now a general key to the world's locked archives goes well beyond one solved letter, since no success rate across other unsolved ciphers has been published. What remains open is whether historians and other cryptographers will confirm the full reading and the re-dating, and how much of the method the model chose versus the person running it.
[drifted from the evidence:] 🔐 Napolyon dönemine ait ve 217 yıldır çözülemeyen 1809 [drifted from the evidence:] tarihli gizli askeri savaş brifingi, OpenAI'ın GPT-6 Astra [drifted from the evidence:] yapay zeka modeliyle sadece 6 saat içinde deşifre edilerek tarihsel bir gizem aydınlatılmıştır. Bu başarı, yapay zekanın geçmişin kilitli arşivlerini açan güçlü bir dijital anahtar haline geldiğini ve tarih ile kriptoloji araştırmalarında yeni bir dönemin başladığını göstermektedir.
[added by the neutral restatement:] An encrypted military briefing dated 1809, [added by the neutral restatement:] from the Napoleonic period, that had been listed as unsolved for 217 years was deciphered in about six hours using OpenAI's GPT-6 Astra [added by the neutral restatement:] model. The post further asserts that this shows AI is now a general tool for opening historical archives and that a new era in historical and cryptological research has begun.
Red-tinted words in the claim drifted from the evidence. Green-tinted words are what a neutral restatement needs.
The trace / claim to source
- An encrypted military briefing connected to Napoleon and dated 1809 was deciphered using OpenAI's GPT-6 Astra. This is reported by the person who did the work and by named outlets including Live Science, Euronews, Tom's Hardware and Calcalist.
- The letter was previously listed as unsolved on Cryptiana's catalogue of unsolved historical ciphers and had not been published in a full reading before this, so describing it as unread for 217 years matches how the originator and the mainstream coverage describe it.
- The content is genuinely a military briefing: it sets out French, allied and Austrian troop positions shortly before Austria's April 1809 offensive, consistent with Napoleon's written order to Eugène of 16 March 1809.
- The figure of about six hours is the figure given by the person who ran the work, for the model's execution time.
- GPT-6 Astra is a real OpenAI model that existed and was being rolled out at the time of the work.
- The result has been accepted by the maintainer of the catalogue that listed the cipher as unsolved, and a reproducible solution package was published, so this is not an unchecked assertion.
- The post says the briefing was deciphered "in just 6 hours," which matches the model runtime but omits that the person running it describes spending six hours of model time plus a few evenings of his own work, and that 33 values of the cipher were already documented in earlier literature. A reader takes away a cold start from zero in six hours, which is not what the source account describes.
- The post's second sentence turns one successful case into a general conclusion that AI "has become a powerful digital key opening the locked archives of the past." The evidence is a single letter, solved once, with no published success rate across other unsolved ciphers. The source account itself locates the achievement in the combination of transcription, cataloguing, translation, research and cryptanalysis running as one workflow, not in a general archive-opening ability.
- Framing the item as a "mystery unsolved for 217 years" implies the cipher resisted serious attack for two centuries. The source account indicates the main obstacle was access rather than difficulty: the only public image was a single low-resolution 1969 reproduction filed under the wrong year, and the most recent solver note on record said in substance that there was nothing available to work on. The person who solved it stated that the cipher itself was not difficult and that cryptographers could have broken it given the time.
- Demo to product conflation, partial: the post presents the outcome as what the model does, with no human in the picture. The work was carried out by an individual engineer who selected the target, supplied the image, and directed and checked the process over several sessions. The model was the tool, not the agent of the project.
- Whether the full reading is correct in every line. The solution package is public and described by its author as mechanically testable, and the catalogue maintainer is reported to have accepted it, but no refereed publication and no independent end-to-end re-solve by another cryptographer has been located.
- How much of the method was chosen by the model and how much was specified by the operator. Practitioner commentary has raised exactly this question, and the division of labour cannot be settled from the published summaries alone.
- Whether the historical re-dating from 1807 to March 1809 has been reviewed by Napoleonic historians rather than by cryptography specialists. No historian's assessment was located.
- Whether the outcome generalises to other unsolved historical ciphers. The author notes the catalogue maintainer is receiving solutions faster than he can record them, which is suggestive, but no systematic measurement of AI success rates on such material was located.
The underlying episode is real and well documented. Carter Church, a staff AI engineer at the cybersecurity company SentinelOne, published a write-up on 18 September 2026 and announced the result publicly on 30 September 2026, stating that he used OpenAI's GPT-6 Astra to read a ciphered letter to General Auguste de Marmont that had gone unread for 217 years. The document is not a letter written personally by Napoleon. It originated at the headquarters of Eugène de Beauharnais, Viceroy of Italy, and was an encrypted troop briefing sent to Marmont on Napoleon's orders. Napoleon's order to Eugène of 16 March 1809 directed that Marmont be told, in cipher, where the French and allied armies stood. The only public image of the ciphertext is a plate printed in 1969 in the French army journal *Revue historique des Armées*, and the catalogue entry carried the wrong date of 1807. The decipherment re-dates the letter to March 1809, roughly two weeks before Austria's invasion. The text runs to about 1,300 cipher units drawn from roughly 155 distinct symbols, in a homophonic substitution system. Church's own account states that only 33 values of the cipher had previously been documented, by the French cryptology historian Daniel Tant, and that the model transcribed the plate, consolidated handwritten symbol variants, wrote a simulated-annealing solver, and checked candidate readings back against the image. He reports re-running the solver with Napoleonic source texts removed from the language model and recovering the same reading. Church's own framing is more limited than the viral framing. He wrote that the impressive part was not the codebreaking but that the model completed the whole multi-modal workflow in about six hours from a single image and a single goal, and he stated separately that the cipher itself was not difficult and that cryptographers could have cracked it given time and energy. He also wrote that he spent six hours of model time plus a few evenings of his own. On verification: Church reports, and multiple outlets repeat, that Satoshi Tomokiyo, who maintains the Cryptiana unsolved-cipher catalogue, reviewed the solution and marked the entry as solved with the date corrected to 1809. Church published a solution package including transcription, key tables and scripts, and says the result is mechanically testable against the ciphertext. On the model: GPT-6 Astra is a real OpenAI model. OpenAI published a launch page for it and CNBC reported the rollout beginning 3 September 2026, so the model named in the post exists and was available at the time of the work.
Complete reasoning
The reply is formatted for pasting into the thread where the claim is circulating.
Ask this case
Answers come only from the case file above; nothing is added.
Was a real Napoleonic cipher actually solved using AI?
Yes. An encrypted 1809 military briefing sent to General Marmont on Napoleon's orders was deciphered using OpenAI's GPT-6 Astra, according to the engineer who did the work and multiple news outlets that covered it.
Did the AI really crack it in just 6 hours from scratch?
The six hours refers only to the model's runtime. The engineer also spent several evenings on the project himself, and a third of the cipher's values had already been documented by an earlier researcher, so it was not a cold start.
Was this cipher truly unsolvable for 217 years because it was so hard?
Not according to the person who solved it. He said the cipher itself was not difficult and that human cryptographers could have broken it given time. The main obstacle was that the only available image was a single blurry 1969 reproduction filed under the wrong year.
Does this mean AI can now unlock any historical archive?
No. That claim goes beyond what the case supports. This is one solved letter, and no success rate across other unsolved ciphers has been published.
Has anyone independently confirmed the solution is correct?
The maintainer of the catalogue that had listed the cipher as unsolved reviewed it and marked it solved, and a reproducible solution package was published. But the investigation did not establish whether a refereed publication or an independent end-to-end re-solve by another cryptographer has taken place.