AI Cracks 217-Year-Old Napoleonic Cipher in 6 Hours,Revealing PreWar Deployments

American AI engineer Carter Church announced on September 30 that he used OpenAI's GPT-6 Astra to fully decipher a 217-year-old Napoleonic-era military cipher in about six hours, completing the workflow from image transcription to cryptanalysis. The cipher, sent in 1809 to French General Auguste de Marmont from Eugène de Beauharnais's headquarters under Napoleon's orders, had been listed as unsolved on the historical cryptography site Cryptiana; its only public complete image was a low-resolution scan from a 1969 French military history journal. Cryptiana maintainer Satoshi Tomokiyo reviewed the solution, marked the entry solved, and corrected the letter's date from 1807 to 1809, with Church re-dating it to late March 1809, roughly two weeks before Austria attacked.

AI Cracks 217-Year-Old Napoleonic Cipher in Six Hours, Revealing Pre-War Troop Deployments A 217-year-old encrypted Napoleonic-era letter has been fully deciphered for the first time with the help of artificial intelligence. American AI engineer Carter Church announced on social media on September 30 that he had used OpenAI's latest model, GPT-6 Astra, to crack the military cipher sent in 1809 to French General Auguste de Marmont. The model's execution time was only about six hours, completing the entire workflow from image transcription to cryptanalysis. The cipher had long been listed as unsolved on Cryptiana, a website dedicated to historical cryptography. The only publicly available complete image was a low-resolution scan published in a 1969 issue of a French military history journal. Over the decades, multiple researchers had attempted to reconstruct its contents, but no one had ever published a complete decipherment. Church published his full research record on September 18, including the transcribed text, cipher tables, and programs that can reproduce the decryption results. According to Church, Satoshi Tomokiyo, the Tokyo-based historical cryptography researcher who maintains Cryptiana, reviewed his solution, marked the entry as "solved," and corrected a previously mislabeled date, changing it from 1807 to 1809. The Letter's Origins and Encryption Method To be precise, this was not a letter written in Napoleon's own hand and sent directly. The cipher originated from the headquarters of Eugène de Beauharnais, Viceroy of Italy, and was an encrypted military briefing relayed to Marmont under Napoleon's orders. In the spring of 1809, war clouds gathered over continental Europe. Austria was preparing to go to war against France again, while a large portion of French forces remained tied down in Spain. Marmont was commanding French troops in Dalmatia, and if Austria launched an attack, he could quickly find himself isolated. On March 16, Napoleon issued orders to Eugène, instructing him to inform Marmont of the positions of the main French and allied forces, specifically requiring that the message be "written in cipher" and delivered by a reliable officer. Based on the decrypted troop numbers, general's titles, and unit positions, Church re-dated the letter to late March 1809—approximately two weeks before Austrian forces launched their attack. The letter's condition was extremely poor. The original image measured only 1202×1836 pixels. The first line could still be read normally in French: "General Marmont, you should have received my letters of the 8th, 14th, and 20th of this month..." followed by 24 lines of numbers, letters, and hand-drawn symbols. Even more challenging, the corresponding complete cipher key had not survived. The cipher was not a simple one-to-one substitution. Church counted approximately 1,300 cipher units in the entire letter, involving 155 distinct symbols. Many high-frequency letters had multiple different written forms, and some symbols represented entire words. This "homophonic substitution cipher" deliberately scrambled letter frequencies, rendering traditional frequency analysis far less effective. Previously, French cipher historian Daniel Tant had identified 33 letter correspondences, explaining 435 cipher units—about one-third of the text. The remaining 865 units still required inference. Another researcher, Daniel Bourdeau, had also attempted to tackle the problem on September 15 but had stalled due to the lack of a complete ciphertext transcription. Church subsequently discovered that the original 1969 journal had been digitized by the French academic platform Persée—the necessary material had been publicly available all along, just hidden in an old scan that was difficult to work with directly. The AI's Decipherment Path Church's day job is related to AI agents. According to his profile, he is currently an AI engineering staff member at the cybersecurity company SentinelOne, with a long-term focus on security operations, incident response, and automation—but with no prior experience in historical cryptanalysis. His approach to the letter resembled a complete agent workflow. First came transcription: Astra sliced the low-resolution scan into line-by-line images and identified handwritten symbols one by one. It initially recorded 1,320 units and 175 provisional symbols, then merged different handwriting variants of the same symbol through comparison, ultimately arriving at 1,300 cipher units and 155 symbols. Next, Astra used the 33 letter values previously identified by Tant as anchors, while also leveraging content written in plain letters—such as "CONSEQUENT"—to expand the known portion. For the remaining symbols, the model wrote a solver program based on a simulated annealing algorithm, then used 19th-century French corpora to judge whether different candidate outputs were natural. The materials used for statistical language features included works by Victor Hugo and Alexandre Dumas, as well as Marmont's own memoirs. During analysis, Astra discovered that treating all symbols as individual letters could not explain the text, so it further classified some symbols as representing complete words. Ultimately, 29 symbols were identified as whole-word symbols, corresponding to common French words such as de , que , les , vous , and général . The final step was verification against the original image. Each time a new symbol mapping was proposed, the model checked all other occurrences of that symbol throughout the letter to see whether they produced plausible results. Church also had the solver re-run after removing Marmont's memoirs and Napoleonic-era texts from the corpus—the same core content was still recovered. It should be noted that five symbols in the solution remain ambiguous for example, "Your Majesty" versus "the Emperor" , but none of these change the letter's meaning. The entire decipherment path can be summarized as follows: Decrypted Content and Historical Corroboration The decrypted content is a remarkably detailed pre-war troop briefing. The letter mentions, in sequence, approximately 40,000 Bavarian troops stationed around Munich and Passau, approximately 30,000 Polish troops positioned near the Vistula River threatening Kraków, and Saxon forces near Dresden. The positions and strengths of French forces under generals Davout, Masséna, and Oudinot are also listed one by one. The letter then assesses that Austria would face attacks from multiple directions and instructs Marmont to act aggressively. Another source supplements more complete troop data: approximately 80,000 French troops stationed around Bayreuth, approximately 60,000 French troops near Ulm and Donauwörth, and roughly 40,000 men under General Oudinot positioned around Augsburg and the Lech River. The letter also mentions that Russian forces were moving toward Austria. The letter uses the customary titles of the era, which readers may not immediately recognize. The table below lists each unit in the order they appear in the letter: | Title Used in Letter | Actual Commander | Strength | Position | |---|---|---|---| | Duke of Danzig | Lefebvre | ~40,000 Bavarian troops | Munich and Passau | | Prince Poniatowski | Poniatowski | ~30,000 Polish troops | Along the Vistula, threatening Kraków | | Prince of Pontecorvo | Bernadotte | — | In front of Dresden | | Duke of Auerstädt | Davout | ~80,000 French troops | Bayreuth | | Duke of Rivoli | Masséna | ~60,000 French troops | Ulm and Donauwörth | | General Oudinot | Oudinot | ~40,000 French troops elite corps | Augsburg and the Lech River area | The order and numbers of these forces correspond almost item-by-item with Napoleon's March 16 orders to Eugène, serving as important external evidence for the decryption's validity. However, at least one detail in the letter was not true. "The Russians are advancing" was a deliberately planted message: on March 21, Napoleon wrote from Malmaison instructing Eugène to "use every means—newspaper articles or other channels—to instill the idea that Russian forces are marching toward Austria." Three days later, he complained to his ambassador in St. Petersburg: "You have not told me where the Russians are. If Russia does not send troops, I will have to deal with Austria alone." In reality, although the Tsar had ostensibly agreed to send troops under the terms of the Erfurt meeting, he privately assured Vienna that he would avoid all conflict. Russian forces did not cross the border into Galicia until June 3—by which time the outcome of the war was already decided. The letter effectively served as a pre-war forecast, and subsequent events largely unfolded along its lines. On April 10, Austrian forces crossed the Inn River without a declaration of war. Archduke John led his forces into Italy from the directions of Villach and Laibach—both named in the letter—while Jelačić's division advanced from Salzburg, and another Austrian force moved through Lienz toward Innsbruck, triggering the Tyrolean uprising. Nearly every enemy assembly point marked in the letter materialized. The estimates of friendly forces, however, proved optimistic: the letter credited Poniatowski with 30,000 Polish troops, when in reality he had about 17,000. He fought only to a draw at the Battle of Raszyn on April 19, and Warsaw subsequently fell. Marmont himself, meanwhile, saw the campaign through from start to finish: he counterattacked successfully in May, captured Laibach—the first enemy stronghold listed in the letter—on June 3, and arrived in time for the Battle of Wagram in July. There is also an intriguing historical detail. In the 1865 published collection of Napoleon's correspondence, the relevant order states that Marmont should not be intimidated by "a handful of..."—followed by an ellipsis. In the decrypted receiving-end cipher, the corresponding passage reads "quelques troupes ou un rassemblement de canailles," roughly meaning "a few troops, or a rabble of scoundrels." However, since this cipher was not a verbatim transcription of Napoleon's original order, it cannot yet be confirmed that this phrasing was Napoleon's own words—only that the message Marmont ultimately received contained it. A String of AI Decipherment Cases Church's decision to attempt cracking the Marmont cipher was prompted by a similar AI decipherment case he had seen earlier. Developer Carter Leffen had used GPT-6 Astra to process a 1941 German Army Enigma telegram. That 82-letter ciphertext, designated MVUEH, had remained unsolved since its public release in 2005. Astra independently selected its target, searched relevant archives, used the place name "ROSENOW ROSENOW" from another already-decrypted telegram sent the same day as a known-plaintext clue, wrote Python and C++ programs to simulate the Enigma machine and cryptanalytic devices, and ultimately found the correct key. Cryptography researcher Frode Weierud later reviewed and confirmed the plaintext and key. On September 20, cryptography researcher Jack Willis used Anthropic's Claude Opus 5 to crack another 1941 German Army Enigma ciphertext, FMNGI. Unlike the previous case, Willis provided Claude with cryptanalytic tools written in Go and relevant historical materials in advance. The final key search was completed on an Apple M2 device in 13 minutes and 28 seconds—though this was only the duration of the final search round, not including the time spent organizing materials and identifying clues beforehand. Another case that drew attention involved a 1918 German ADFGVX radio ciphertext. GPT-6 Astra found a set of keys already published in historical sources and cross-checked the decrypted time and location against the navigation records of the British warship HMS Canterbury. However, this case still has unresolved issues: the key used in the decryption was not scheduled to enter service until December 9, 1918, according to known records, while the message was sent on November 27—a chronological mismatch—and no independent reproduction has yet been published. Taken together, the four cases differ in the degree of human involvement and the nature of external verification: | Case | Era & Cipher | Model Used | Human Involvement | External Verification | |---|---|---|---|---| | Letter to Marmont | 1809, homophonic substitution cipher | GPT-6 Astra | Only task assignment; transcription, decipherment, and tool-writing all done by model | Reviewed by unsolved-list maintainer and marked "solved" | | Enigma telegram MVUEH | 1941, Enigma | GPT-6 Astra | Only asked to attempt unsolved-list items; model selected target and wrote Python & C++ tools | Plaintext nearly identical to same-day decrypted SIPVX | | Enigma telegram FMNGI | 1941, Enigma | Claude Opus 5 | Human provided Go-language analysis tools and name clues | Checked against archived transmission copy | | ADFGVX message | 1918, ADFGVX | GPT-6 Astra | Target specified by researcher | Questionable; key activation date later than message date | As such, these "AI cracks historical cipher" cases currently have varying degrees of external verification. The Real Significance The "six hours" in "AI cracks 217-year-old cipher in six hours" primarily refers to GPT-6 Astra's model execution time. After completing the decipherment, Church spent several additional evenings organizing materials, verifying historical context, and producing files that others could use to reproduce the results. No new cryptographic principles emerged from the process. What is truly noteworthy is that a task that previously required coordinated efforts across image transcription, historical document retrieval, linguistic analysis, cryptanalysis, and software development was completed and connected end-to-end by Astra in a continuous workflow. The Marmont cipher's broader significance lies in demonstrating how AI lowers the research cost of such historical problems. The scan had long been public, partial cipher tables had been compiled by others, and Napoleon's correspondence, Marmont's memoirs, and 19th-century French corpora were all accessible. What had previously consumed the most time was locating these scattered materials and then completing the transcription, cataloging, programming, and iterative verification. In his retrospective, Church also noted that a patient cryptography expert could likely have solved this letter decades ago—but few people were willing to invest weeks of time on a low-resolution scan buried in a military history journal. Now, models compress this kind of work into a much shorter timeframe. This shift may extend beyond cryptography. Historical archives, ancient scripts, manuscripts, old maps, and even scientific data scattered across different databases all contain similar problems: the materials and methods already exist, but the investment cost was previously too high. As AI increasingly takes on document retrieval, image recognition, programming, and result verification, the people who tackle these "unsolved mysteries" in the future may increasingly come from outside specialized fields. Add preferred source https://www.google.com/preferences/source?q=finance.biggo.com Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.