Amateurs using AI are prying open historical ciphers once thought to be out of reach
A wave of non-experts is now using large language models to crack historical codes, and the pace is fast enough to alarm the specialists who are being asked to verify the results. In one recent case, a codebreaker with no cryptography training asked an AI model to work on a German army message encrypted 85 years ago, nudged it to keep thinking, and woke up to a proposed solution. Another hobbyist used the same general approach to decode a 1809 letter tied to Napoleon’s army. Both examples point to the same shift: language models are not just generating text, they are increasingly being pressed into service as computational assistants for old-fashioned cryptanalysis.
The interest is understandable. Historical ciphers often reward persistence, pattern matching and a willingness to test thousands or millions of possibilities, all tasks that suit machine assistance. Large language models can write code, search archival sources, transcribe images of encoded text and propose candidate keys in one workflow. They also introduce a new bottleneck. Cryptography experts are being flooded with requests to validate supposed breakthroughs, many of which collapse under scrutiny.
“There are so many false positives that are coming in from people who don’t really understand cryptanalysis,” said Elonka Dunin, a retired game developer in the Washington DC area who has written books on cryptography.
Enigma settings, a crib and more than four billion tries
Carter Leffen, a product development coach at Bloomberg in New York City, used GPT-6 Astra to work through one of the unsolved messages listed by Crypto Cellar Research, a site run by independent cryptographer Frode Weierud. Leffen selected a German army message from 10 July 1941, a date that also mattered because another Enigma message from the same day had already been decoded in 2017. Enigma, the machine used by the German armed forces in the Second World War, encrypted letters through rotors and plugboard settings that could generate up to 103 sextillion combinations.
The AI model focused on the short length of the message — 82 characters — and on the possibility that it shared text with the previously solved message. In cryptanalysis, a known or suspected fragment of plaintext is called a crib, and it can narrow the search enough to make a brute-force attack tractable. GPT-6 Astra wrote code to simulate an Enigma machine and searched for settings that would both reproduce the crib and make the text intelligible. Over ten hours, it tested more than four billion settings.
The result was not just a lucky guess. The message was encrypted with settings different from those used in the already-deciphered message from the same day, an unusual detail that likely kept the code intact for so long. For a machine designed to scatter letters through a vast keyspace, that kind of difference can matter enormously; it removes the shortcut that a related message might otherwise provide. A single change in daily settings can preserve secrecy for decades.
Simulated annealing and a 19th-century troop movement
The second case followed a similar logic but a different historical setting. Carter Church, a staff engineer at cybersecurity company SentinelOne in Mountain View, California, turned GPT-6 Astra toward an unsolved cipher sent in 1809 by an Italian official to one of French emperor Napoleon’s generals. The message had been posted on Cryptiana, a site assembled by Japanese patent engineer and hobby cryptographer Satoshi Tomokiyo.
Church provided the model with an image of the encrypted letter. The system did more than simply guess at the answer. It found a partial key among records on a French cryptography website, then applied an incremental method known as simulated annealing, in which a key is adjusted step by step and retained when the output becomes more coherent. In about six hours, GPT-6 Astra reached a successful key. The decoded message described troop locations.
Church’s assessment was pragmatic. The message probably could have been solved without AI, but the model cut down the time and effort by handling several labor-intensive steps in sequence: transcribing symbols, finding a partial key and deciding how to apply it. That combination is exactly where these systems are strongest. They are less reliable as final arbiters than as assistants that can compress the first hours of a search into a few clicks and a long computational run.
False positives are becoming the real workload
The flood of amateur successes has created a second-order problem for cryptographers. As more hobbyists try old codes with AI help, more incorrect solutions reach the inboxes of people capable of judging them. Those validations are not simple courtesy checks; they require careful review of whether a proposed plaintext fits the historical context, whether the key search was logically sound and whether the apparent solution survives independent testing.
That burden is especially heavy because large language models can produce outputs that look convincing even when they are wrong. In cryptography, a plausible sentence is not enough. A decoding has to fit the method, the key structure, the source material and the historical record. AI systems can help generate hypotheses, but they can also amplify overconfidence, especially when used by people who do not know the difference between a partial match and a real break.
The broader significance reaches beyond codebreaking hobbyists. Historical cryptanalysis has always lived at the intersection of computation, patience and expert judgment. Large language models now lower the barrier to entry, but they do not erase the need for domain knowledge. If anything, they increase it. Once a machine can produce a polished answer in hours, the human task shifts toward deciding which answers deserve belief and which are merely artifacts of a very convincing search process.
Why these machines are effective, and where they still fail
LLMs work well in this setting because they can orchestrate several tasks at once. They can identify likely sources, retrieve background material, write code to model an encryption system and iterate through candidate keys far faster than a person working by hand. In both cases described here, the models were not simply “solving” a cipher in the abstract; they were building a workflow around the cipher. That workflow matters because old codes often fail at the margins rather than in the center. One historical clue, one partial key or one related message can change the odds dramatically.
Still, the technology is not a substitute for cryptanalytic reasoning. A system that can search billions of combinations still needs a principled way to judge which output is meaningful. A decoded message about troop movements or a wartime German communication may look obvious after the fact, but the hard part is the path from raw symbols to a defensible solution. When that path is handed to a model, the final check remains stubbornly human.
For now, the surge in AI-assisted codebreaking is producing both genuine breakthroughs and a backlog of mistaken claims. The backlog may grow faster than the breakthroughs unless the people submitting solutions learn to separate novelty from proof.