AI cracked a centuries-old cipher? We fact-checked ten celebrated cases

What this is about

Since mid-September 2026, a wave of reports about cracked historical ciphers has swept the internet. Almost every day, amateur researchers announce that ChatGPT (GPT-6 Astra) or Claude has solved a cipher that experts had failed to break for decades. Cryptographer Cécile Pierrot (Inria) and historian Camille Desenclos (Université de Picardie) counted 200 to 300 such claims within a few weeks, writing in The Conversation. Their verdict:

"Almost none of the results have been verified, and several are flawed or already known."

We've enthusiastically reported some of these cases too, most recently Napoleon's letter to Marshal Marmont. So it's time to take stock. We checked ten well-known cases against four questions:

1. Is the solution unique and reproducible?
2. Are the code, transcription and key publicly available?
3. Has an independent expert confirmed it?
4. Was the content already known?

The scorecard: ten cases, four criteria

Genuine, but not much new

Napoleon to Marmont. We reported this one ourselves without context. In terms of craftsmanship, it's the cleanest case. On carter.church, Church publishes a package with the transcription, the full key and scripts (verify.py, reproduce.py). Anyone can use it to reproduce the reading offline. Tomokiyo checked the solution and set the Cryptiana entry to "solved." Five symbols remain uncertain because they appear only once.

We got two things wrong back then. First, the starting point: Daniel Tant's previously known partial table already supplied 33 symbol values. They cover 435 of the roughly 1,300 units, a third of the letter. Church also mentions "about six hours of compute plus personal evenings." So this was nowhere near "a single prompt." Second, the content: the letter came from the headquarters of Eugène de Beauharnais and wasn't written by Napoleon himself. Its list of forces matches the order of March 16, 1809, which has appeared in Napoleon's printed correspondence since 1865. What's actually new is essentially half a sentence missing from the printed version ("a pack of rabble"). That's why Pierrot and Desenclos say the letter brought "nothing new" to research.

The Enigma messages. For MVUEH, dated July 10, 1941, Weierud writes on Crypto Cellar that GPT-6 Astra worked "completely independently." The model picked the message itself, recognized its relationship to the already broken message SIPVX and used the repeated place name ROSENOW as a crib (a guessed piece of plaintext). It then wrote an Enigma simulator and a "Bombe" (the search machine the British used to find Enigma keys) in Python and C++. According to Weierud, the key and plaintext were immediately recognizable as correct. In substance, though, MVUEH repeats SIPVX almost word for word. It's also unclear where the model learned about the "private collection" and the Bundesarchiv shelf marks that show up in its logs.

For FMNGI, dated July 31, 1941, Jack Willis gave Claude Opus 5 ready-made analysis software, source material and a direction. The actual key search took 13 minutes and 28 seconds on an Apple M2, according to Weierud. The plaintext reads: "KOLONNE VON X KORPSNACHSCHUB ZURUEQ X HARTJENSTEIN X HARTJENSTEIN" (roughly: column of corps supply back, Hartjenstein). Of the collection's roughly 1,000 messages, only seven Enigma messages now remain unbroken.

Baltiysk. Anton Zaytsev solved the Baltiysk bottle message with Claude, and his solution is documented on GitHub with statistical tests. Klaus Schmeh calls it "remarkable". Yet hundreds of grid and transposition methods turned up no key. The solution came another way: the letter frequencies were matched against song collections. The result is the East German youth song "Einheitslied," spread across six scrambled 13×13 squares. Schmeh himself writes that it's unclear whether any systematic method is behind it at all.

New, but with open questions

Koehler 1944 and Wood 1950. Both cases rely on a book key, where a passage from a book serves as the key. For Koehler, it's a page from the Dutch prayer book "Offeraltaar des gebeds" (1893). For Wood, it's the Lord's Prayer from the Luther Bible. Our take: solutions like these are hard to fake. If a specific, verifiable book passage yields meaningful German, that's strong evidence. With Koehler, though, there's a dispute over credit. Czech amateur researcher "Viktor" had already reported the prayer book to Schmeh on September 19. Schmeh doesn't describe recalculating it himself.

ADFGVX 1918. The plaintext reports an English cruiser in Sevastopol and an Allied squadron "on the 26th." That matches the logbook of HMS Canterbury. But the keyword TRUPPENVERSCHIEBUNG was already listed in J. Rives Childs' standard work on German military ciphers. According to Childs, it only came into use on December 9, 1918, while the message is dated November 27. So the model tried a known key. Schmeh merely passes on the claim, and according to 20 Minuten, nobody has cross-checked it yet.

What doesn't hold up, or can't

Urquhart's Distich. On August 31, Vals AI announced that Claude Fable 5.1 had solved the 64 numbers in 44 minutes. The result: "O GOD UPHOLD KING CHARLS THE SECOND …". The very next day, the Reticuli group published a counter-check with scripts. In the British Library copy, the number sequence doesn't appear in the book at all. In ten places, not a single word in the relevant section starts with the required letter. Out of 65 counting schemes tested, the best ones hit 8 of 64 characters, which is chance level. The picture is muddled, though. In Bruce Schneier's comment section, one reader writes that the method is reproducible with the 1834 edition, and that a copy in Edinburgh contains the distich. Zodiac solver David Oranchak thinks the solutions are probably correct. Schneier himself wrote in an update that he's "not sure". Vals AI hasn't released its internal files.

Memorandum from 1560. According to Pierrot and Desenclos, the transcription published on a blog contains errors that aren't flagged. Parts of the plaintext are also missing or were altered without mention. A decryption can never be better than the transcription it's based on.

Zodiac Z13. Here, math decides. The unicity distance is the minimum length at which a ciphertext allows only one meaningful solution. Claude Shannon described it in 1949. Below that length, many different plaintexts fit equally well. Even for a simple substitution cipher, it sits at around 28 characters. The nibzard/zodiac project estimates it at 20 to 30 characters for homophonic ciphers, in which common letters are represented by several interchangeable symbols. Z13 has 13 characters. The project's solver found 52 different readings that score almost equally well, and more than a thousand slightly further behind. According to the SF Standard, Alex Baber arrived at his "Marvin Merrill" solution by filtering 71 million names. Former NSA cryptographer Ed Giorgio only confirmed the "internal consistency" of the method, not the name. In our view, anyone announcing a unique solution to Z13 has simply picked one of many fitting readings.

Why most successes are grunt work

The solid cases follow a pattern. First comes the transcription: copying the original into characters. Then come classic methods that have been around for decades. One of them is simulated annealing: a search program tries millions of mappings and initially accepts worse ones too, so it doesn't get stuck in a dead end. Add crib tests and trying out known keys. The language model writes the code, searches archives and keeps at it. Church puts it himself: "What AI mainly brings is persistence." According to one Schneier commenter, Church's write-up even says a patient specialist could have cracked the letter back in 1970.

The numbers back this up. Daniel Bourdeau runs the open directory Cyphersolver with Claude, and George Lasry has spot-checked it. Of the 603 cases processed there, 233 keys were actually recovered through cryptanalysis (39%). In 267 cases (44%), a known key was applied or an existing decryption was read, and 103 remain unsolved. This directory, notably, discloses its categories. That's exactly what most one-off social media announcements lack.

Our verdict

The wave isn't a hoax. At least five of the ten cases survive strict scrutiny: Marmont, both Enigma messages, Baltiysk and, very likely, Koehler and Wood. What's mainly new is the speed. New historical knowledge emerges less often than the headlines promise. Marmont and MVUEH essentially repeat known texts, and the 1918 radio message was decrypted with a known key. The Distich case is open, the 1560 memorandum isn't reliable, and Z13 can't be uniquely solved in principle.

The biggest problem is verification. Take the two working days per case from the Inria critique, and 200 to 300 claims add up to 400 to 600 working days. By our math, that's two to three person-years, which a handful of experts are expected to handle on the side. How many submissions Schmeh, Lasry or Inria have confirmed or rejected hasn't been published. Until then, judge every "AI decryption" by two things: Is there data that lets others reproduce the solution? And is there the name of an expert who checked it? If both are missing, the announcement is just a claim for now. In its "Back in Time" project, Inria is working on a free, specialized transcription model designed to deliver verifiable results.