GPT-6 Astra Cracks a 217-Year-Old Napoleonic Cipher From a Single Blurry Image

The administrator of Cryptiana — the internet's primary repository of unsolved historical ciphers — reviewed Carter Church's methodology and declared the 1809 Napoleon-to-Marmont dispatch officially closed. The letter had appeared on Cryptiana's unsolved list for decades; every cryptographer who examined it found no workable basis for attack. Church, a senior AI engineer at cybersecurity firm SentinelOne with no cryptography training, handed GPT-6 Astra a single low-resolution black-and-white scan in September 2026 and let the model run. Six hours later, he had a complete cipher key, a full French plaintext, and the recovery of a passage Napoleon's 19th-century editors had redacted from the historical record.

The story went viral on September 30 when Church posted his methodology on X.

What the 1809 Cipher Was and Why Experts Left It Alone

On March 16, 1809, Napoleon ordered a top-secret dispatch to Marshal Auguste de Marmont, who commanded 13,000 French troops across the Adriatic as Austria prepared to attack. The letter had to travel through hostile territory, so Napoleon specified it be encoded by "an extremely intelligent officer." The resulting cipher spans 24 lines of two-digit numbers, stray letters, and invented symbols — with only the opening sentence, a standard greeting, written in plain French.

The cipher key was lost in the wars that followed. The only surviving copy of the letter is a 1969 photocopy from a French military magazine, reproduced at 1,202 × 1,836 pixels — an average of 19 pixels per cipher symbol. Cryptanalyst Daniel Bourdeau examined it weeks before Church and gave up: the image quality was too poor, the key was gone, and no cryptographer could justify the interdisciplinary effort.

The cipher's design compounded the difficulty. It is homophonic: high-frequency French letters like e, a, and s each map to five to eight different symbols, deliberately flattening the ciphertext's frequency distribution and defeating the standard frequency-analysis approaches that modern cryptanalysts apply first. With 1,300 symbols across 155 distinct graphemes and no surviving key, the letter had no conventional entry point.

How GPT-6 Astra Solved It: Five Steps Over Six Hours

Church's prompt was minimal: he supplied the scan and asked the model to decode it. What followed was an uninterrupted autonomous session.

Step 1 — Visual transcription. GPT-6 Astra segmented the image row by row, isolating individual cipher symbols and normalizing scribal variation. The model identified that symbols appearing in slightly different forms were the same grapheme written with the natural inconsistency of 19th-century handwriting. It classified 1,300 occurrences into 155 canonical symbol types — a task that Church notes would take human paleographers several months using manual comparison.

Step 2 — Anchoring on known values. Astra retrieved a partial cipher table assembled by French cryptology historian Daniel Tant of ARCSI, the French cryptology reserve organization, which established 33 letter-to-symbol correspondences. It also detected "CONSEQUENT" — the French word for "therefore" — spelled out in ordinary Latin letters in line 14, a scribal exception that served as a known-plaintext crib. These two anchors gave the model approximately one-third of the key and a foothold into the letter's language.

Step 3 — Writing and running a simulated annealing solver. With 865 unknown symbol-to-letter mappings remaining, frequency analysis was useless against a homophonic cipher. Astra wrote its own Python solver using simulated annealing — a probabilistic optimization algorithm that searches for optimal solutions by accepting imperfect moves early in the process and becoming increasingly selective as it converges. The fitness function evaluated candidate decryptions against French n-gram statistics: how often each proposed letter sequence appears in authentic 19th-century French text. For that corpus, Astra drew on the works of Victor Hugo and Alexandre Dumas, plus Marmont's own published memoirs — a choice that made the n-gram model unusually well-calibrated to the exact register of the letter.

Step 4 — Discovering whole-word symbols. Partway through the solve, the model detected that certain symbols produced grammatically impossible letter combinations when treated as individual characters. It tested an alternative hypothesis: some symbols represent complete words rather than individual letters. That hypothesis was correct. Twenty-nine symbols in the cipher correspond to common French military and grammatical terms — de, que, les, vous, général — a shorthand convention that makes the cipher faster to write and even harder to crack without knowing the convention exists.

Step 5 — Cross-validation. Before producing a final output, Astra re-encoded the solved plaintext using the derived key and compared it symbol by symbol against the original scan, confirming internal consistency across the full letter.

The output was a 144-entry key table — 115 letter-symbols and 29 word-symbols — along with a complete French plaintext, published with verification scripts enabling independent reproduction.

Read more: GPT-6 Astra launches as OpenAI's first Critical-rated cybersecurity AI

What the Letter Said, and What Napoleon's Editors Hid

The decoded dispatch is a pre-battle intelligence briefing. Napoleon reported the positions and troop strengths of his marshals — Davout, Masséna, and others — and closed with a characteristically contemptuous assessment of the Austrian forces massing against him. The letter describes Austria as "consequently hastening toward her ruin." The French would win the Battle of Wagram that July, and Napoleon would promote Marmont to marshal.

The more significant historical discovery was in a passage that historians had long known was incomplete. In the 1865 publication of the Correspondance de Napoléon Ier — the official 32-volume collection of Napoleon's letters — one sentence in the Marmont dispatch appeared with an abrupt ellipsis: "Marmont must not be frightened by a handful of... (par une poignée de . . . . . . . .)" Whether the original manuscript was damaged, or whether 19th-century editors under the Second Empire found the phrase politically inconvenient, was unknown.

The deciphered letter fills the gap. According to Church's decode, the cipher letter carries a paraphrase of the same passage from Marmont's receiving end, which reads: a few troops, or a gathering of rabble (quelques troupes ou un rassemblement de canailles). Church notes this relay is a paraphrase rather than Napoleon's exact words, but it is the first direct evidence of what the ellipsis was concealing — language that Second Empire editors may well have wanted to suppress given the diplomatic climate of the 1860s. The phrase has not yet been formally reviewed by Napoleonic historians, and that authentication step remains pending. The Fondation Napoléon has noted that the modern critical edition of the Correspondance générale re-edits 1809 letters from originals and would be the place to check whether the missing words survive elsewhere.

The Expertise Cost Problem That AI Just Solved

Church's own summary is pointed: "The cipher wasn't mathematically hard. Cryptographers could have solved it years ago. The problem was that it required a historian, a paleographer, a cryptanalyst, and a programmer working together on something nobody thought was worth funding."

The Cryptiana database maintains hundreds of unsolved historical ciphers; the DECODE database at Uppsala University tracks thousands of encrypted historical sources. These persist not because they are computationally intractable but because the expertise required crosses disciplines that do not regularly collaborate, and the research value of any individual cipher rarely justifies the cost.

GPT-6 Astra collapsed that cost. Its 1,050,000-token context window can hold extensive historical reference material alongside the active problem; its multimodal pipeline processes degraded document images; its code generation executes domain-specific solvers; and its long-horizon agentic design runs a multi-step research plan without human steering. The six-hour solve is not primarily about raw intelligence — it is about a coordinated capability set that previously required four specialists and months of calendar time.

The same week, two further historical ciphers fell to Astra: a World War II Enigma variant decoded in approximately ten hours by Carter Leffen, a product developer at Bloomberg, and a German World War I radio cipher solved using an independently discovered key. The pattern suggests the Marmont letter is not an isolated demonstration.

GPT-6 Astra's Capabilities and Where It Sits Competitively

GPT-6 Astra, released by OpenAI on September 3, 2026, supports text and image inputs across a 1,050,000-token context window at $10 per million input tokens and $50 per million output tokens. Independent benchmark organization Artificial Analysis places it at 61 on its Intelligence Index — five points below Anthropic's Claude Fable 5.1 at 66, based on comparable effort-level testing.

The Napoleon cipher demonstration is a user-generated capability test rather than an OpenAI-designed benchmark. The task was image parsing of a degraded historical document, not a formatted OCR problem; code generation for a domain-specific solver, not a standard algorithm; and corpus retrieval calibrated to 19th-century French literary text. Whether Claude Fable 5.1 or other contemporary frontier models could replicate the result has not been tested. The Marmont cipher provides a concrete, reproducible benchmark for the specific capability cluster — sustained long-horizon planning combining multimodal vision, code execution, and domain corpus retrieval — that the solve required.

Read more: GPT-6 Astra breaks a solo developer's multimodal benchmark built by coding agents

What Comes Next for Archival AI

Satoshi Tomokiyo's confirmation changes the terms of what archival institutions should be planning. Every major historical archive holds material that has resisted human scholars not because it is impossible to analyze, but because the cost in expert time exceeded any plausible research budget. The calculus is now different.

The more immediate question is whether the recovered passage — a gathering of rabble (un rassemblement de canailles) — will survive scrutiny from Napoleonic historians examining the original 1865 publication gap. That authentication step is the one piece of the puzzle Church's solve cannot complete automatically, because it requires contextual judgment about 19th-century editorial practice that sits outside any cipher table. Church's verification scripts are publicly available, and the methodology is reproducible; the historical verdict is the remaining open thread.

Beyond Napoleon, the Cryptiana and DECODE unsolved-cipher lists have not changed materially in years. They are about to.