AI's Achievement: World War I Code Cracked After 107 Years

AI's Achievement: World War I Code Cracked After 107 Years

170 Characters Unread for 107 Years

Sixteen days after the war ended, a radio message was recorded by the Germans.

The transmission date was November 27, 1918. It used the "ADFGVX cipher," operated by the German military in the late stages of World War I. The message consisted of 170 characters of cipher symbols, and even as many similar communications were deciphered, this one long resisted decryption.

Over 107 years later, in September 2026, an author under the pen name prinz reported that they might have used OpenAI's AI model "GPT-6 Astra" to decipher this message for the first time.

Although the restored German text had some missing parts and unnatural spellings, the main idea was clear.

"A British cruiser entered Sevastopol on the 24th. Allied fleet to follow on the 26th."

If it were just a meaningful sentence, there would be room to suspect coincidence or a "plausible wrong answer." The main reason this decryption proposal gained attention is that the two dates mentioned matched records left by the British Navy.


The Real Cruiser Behind the Cipher

According to the log of the British Navy's light cruiser HMS Canterbury, the ship departed Constantinople on November 23, 1918, and was moored to a buoy in Sevastopol harbor at 10:20 a.m. the following day.

Furthermore, the record for November 26 states that "the Allied fleet arrived" at 11:30 a.m.

The text derived by AI included multiple elements such as "British cruiser," "Sevastopol," "24th," "Allied fleet," and "26th." Since these correspond with the ship's log, strong circumstantial evidence was created for the decryption proposal.

However, the date portion in the cipher text is not complete. The restored text contains a section that can be read as "S4STEN," which the proposer interprets as "?4th." Considering the ship's log, the 24th is the most likely, but the number 2 was not clearly derived from the cipher text itself. Possible causes include omissions during transmission, errors in reception or transcription, or damage to the historical materials.

In deciphering historical documents, it is important not to erase this small uncertainty. Instead of concluding "it's confirmed as the 24th because it matches the log," it is more appropriate to express it as "reading it as the 24th aligns very well with independent records."


Why the ADFGVX Cipher Was So Formidable

The ADFGVX cipher converts plaintext into another sequence of symbols using only the six letters A, D, F, G, V, and X, and then rearranges that sequence in a two-step process. One reason these six letters were chosen is that their Morse code forms are easily distinguishable, reducing errors in radio communication.

In the first stage, a 6x6 table is filled with letters and numbers. Each character or number is replaced by two symbols indicating the row and column, so one original character becomes two, like "AD" or "VX."

In the second stage, the intermediate string is written horizontally under a keyword, and the columns are read out according to the order of the keyword's letters arranged alphabetically. This separates the originally paired symbols, making it difficult to find the original text through simple frequency analysis.

For the 170 characters in this case, creating a table with rows of 19 characters to match a 19-character keyword fills 8 rows, leaving 18 characters. Thus, 18 columns have 9 characters each, and one column has 8 characters. While considering these uneven column lengths, the cipher text must be returned to the correct columns.

At first glance, it seems like a mechanical task, but unless the correct keyword, column order, position of the short column, and conversion table correspondences are aligned, the result will not form a meaningful sentence. If any one assumption is incorrect, the result quickly reverts to a meaningless string of symbols.


The Key Word "Troop Movement"

The breakthrough is said to have been the German word "TRUPPENVERSCHIEBUNG," which means "troop movement" or "force movement" in English.

This word itself was not something the AI invented from nothing. It was documented in J. Reeves Childs' history of German military ciphers and was known as a candidate for a key used at the time. The AI appears to have selected clues from existing historical materials, applied them to the structure of the cipher text, and cross-referenced the restored result with external historical records.

Here lies an important point in understanding the significance of this achievement.

This is not a story of "AI instantly destroying an unknown encryption method." The encryption method was known, and candidates for keys were documented. The challenge was selecting promising candidates from scattered information, verifying them in the correct procedure, and supporting the resulting text with independent materials.

The value of AI lies not only in superhuman insights but also in handling numerous candidates, maintaining calculations, judging linguistic validity, and advancing to the next investigation, such as "Can this place name and date be verified in other records?" The ability to connect this series of tasks in a short time is where its value lies.


Why It Hadn't Been Solved Until Now

The hypothesis presented by the proposer suggests that "TRUPPENVERSCHIEBUNG" was considered a key used from December 9, 1918, onward. However, the target message is dated November 27. If previous researchers strictly limited the usage period of the key, the priority to test this word in November communications would have been low.

In other words, the obstacle may have been the assumptions in historical materials, not computational ability.

In cryptography research, lists of known keys and operational dates are powerful clues. However, if there are discrepancies or exceptions in those records, the clues can overly narrow the search range. AI may have reached a seemingly correct sentence by testing candidates beyond those boundaries.

However, it is not known why the start date for using the key is different. Whether the record's date was inaccurate, it was used earlier than planned on-site, or the same word was used in a different system earlier. This requires further historical material research.


Can We Definitively Say "AI Solved It"?

This story requires consideration at three levels.

First, the restored text is meaningful German. Second, its content specifically matches the log of HMS Canterbury. These points strongly support the decryption proposal.

Third, whether this is truly the first decryption in history, with no oversight in the procedure, and superior to other keys or readings, requires independent expert replication. The central evidence currently available is prinz's public article and the reports introducing it. No peer-reviewed papers or formal verification results by cryptographic historians have been confirmed.

Additionally, the claim that "there was no answer in the training data" should be handled cautiously. The proposer claims not to know of past decryption examples, but it is difficult to fully prove the model's learning content from external sources, such as online, books, or unorganized archives. Even if it seems unlikely that a known answer was simply reproduced, it cannot be definitively stated as "absolutely not included."

Therefore, what is important is not the flashy model name but reproducibility. If the used cipher text, key candidates, column restoration rules, conversion table, and each stage up to the final text are published, and third parties can reach the same result, the value of the achievement can be evaluated independently of trust in AI.


Reactions on Social Media and Comment Sections—A Mix of Surprise and Skepticism

This topic spread across X, Hacker News, Substack, and Golem's comment sections, with reactions largely divided into three categories.

 

The most prominent are positive reactions such as "It's fascinating that communication from over 100 years ago connects with the latest AI" and "It's ideal for AI utilization in historical research." Many felt that the connection of separate materials like military history, ciphers, and ship logs demonstrated a use of AI beyond mere text generation. The proposer's Substack also received responses praising the achievement.

On the other hand, there were also calm observations from those knowledgeable about technology. Comments included, "If known key candidates were applied, this is different from breaking an unknown cipher from scratch," and "It should be clarified not only the AI's inference ability but also which materials and candidates humans provided."

On Hacker News, there was discussion on how AI agents excel in problems with clear verification conditions, and how the given prompt and candidate set greatly influence the outcome. In Golem's comment section, points were raised such as "It shouldn't be considered evidence of superintelligence," "How were unknown numbers handled," and "Did operational errors or quirks in cipher procedures at the time affect decryptability?"

Furthermore, there is anticipation for an era where AI reads historical documents. Not only undeciphered ciphers but also hard-to-read handwritten documents, ancient texts with variant characters, and vast ship logs and administrative records, where the biggest constraint was traditionally the time of experts, can accelerate candidate extraction and cross-referencing.

However, reactions on social media are not expert consensus. The current excitement should be seen as heightened interest in public verification rather than confirmation of the achievement.


AI May Change How We Connect Investigations Rather Than "Discoveries"

If the decryption proposal is ultimately endorsed by experts, its significance will not be limited to reading a single military message.

Historical research involves materials of different formats and storage locations, such as cipher tables, lists of keys, old specialized books, original documents, ship logs, and changes in place names. Human researchers can read each deeply, but it takes an enormous amount of time to cross-test related candidates.

AI is not a device that replaces researchers to determine history. Rather, it can become a tool that constantly suggests the next step in the investigation, such as "This key can be tested," "The ship mentioned in this restored text can be confirmed in logs," and "Date omissions can narrow down candidates from other materials."

In that case, the human role does not diminish but becomes more important. It is necessary to verify the provenance of materials, identify parts conveniently supplemented by the model, search for materials that serve as counter-evidence, and judge the degree of certainty to give to conclusions.


The Questions a 107-Year-Old Communication Poses to the Modern Era

The cipher text conveyed the movements of a British cruiser and the Allied fleet. The content itself is brief and not a secret that would change the course of the war. However, the process by which this short report became readable over a century later symbolizes what modern AI is suited for.

AI did not suddenly create truth in a place without answers. It found connections buried between the keys left by humans, the structure of the cipher, the 170-character record, and the British ship's log, and presented them as a verifiable hypothesis.

If third parties reproduce the same procedure in the future and cryptographic historians confirm its validity, this communication will likely be removed from the list of "undeciphered." Conversely, if defects in the procedure or alternative solutions are found, it will become a valuable counterexample to AI's plausible reasoning.

Either conclusion has meaning. What is needed for discoveries in the AI era is not just surprise. It is to publish procedures, return to primary sources, and make it verifiable by others. The most important message conveyed by the cipher that slept for 107 years may lie there.


Source URL

Related Articles