Amateurs are using large language models to crack historical ciphers
A new kind of codebreaking is unfolding outside the traditional cryptography community. People with no formal training in cryptanalysis are feeding historical ciphers into large language models and getting workable solutions, sometimes after only a few hours of model-guided searching.
The pattern is already visible in two recent cases. One involved a German army Enigma message that had remained unread for 85 years. The other involved an 1809 cipher sent by an Italian official to one of Napoleon’s generals. In both instances, the AI did more than suggest possibilities. It helped select the target, transcribe the material, generate code, search for key settings and iterate through candidate decryptions until the text became legible.
The speed is what has startled specialists. So is the fact that people with no prior background in cryptography can now participate in work that once demanded deep technical fluency, patience and access to specialized tools.
GPT-6 Astra handled selection, transcription and search in one workflow
Carter Leffen, a product development coach at Bloomberg in New York City, asked GPT-6 Astra to scour Crypto Cellar Research, a site that lists historical unsolved ciphers. He instructed the model to select an Enigma message and try to decode it. The system chose a message sent on 10 July 1941, likely because it was only 82 characters long and might have been linked to another Enigma message from the same date that had been solved in 2017.
That potential overlap mattered because a previously solved message can supply a crib, a likely fragment of plaintext that narrows the search. In this case, the model wrote code to simulate an Enigma machine and began searching for settings that could both reproduce the known crib and make the remaining message intelligible. Over ten hours, it tested more than four billion settings.
The result was not a lucky guess. Enigma had been built to scramble letters through rotors and plugs in up to 103 sextillion possible ways, so a brute-force-style search still demands careful strategy. Leffen sent the solution to Frode Weierud, the independent cryptographer who runs the website, for confirmation. The decrypted message used settings different from those in the already solved message, even though both were sent on the same day. That kind of inconsistency helps explain why the cipher remained unsolved for so long.
The breakthrough came from combining software generation, historical context and persistence in a single loop.
AI is turning historical cryptanalysis into an iterative search problem
Carter Church, a staff engineer at cybersecurity company SentinelOne in Mountain View, California, took a different route a few days later. He used GPT-6 Astra to tackle an unsolved cipher from 1809, sent by an Italian official to a French general serving under Napoleon. The message had been listed on Cryptiana, a website compiled by Japanese patent engineer and hobby cryptographer Satoshi Tomokiyo.
Church gave the model an image of the message. From there, the system first helped transcribe the symbols, then located a partial key on a French cryptography website, and finally applied simulated annealing to search for a usable decoding key. Simulated annealing works by changing a candidate key step by step and checking whether the output becomes more readable. The model reached a successful key in about six hours.
The decoded message described troop locations.
That kind of result reveals why these systems are proving attractive to hobbyists. A person no longer needs to handle every part of the workflow manually. The model can do the transcription, search the web for supporting fragments, draft code for a decryption engine and keep trying variants until a message becomes coherent. For older ciphers, where the answer may depend on a partial key or a narrow historical clue, the ability to move quickly across several tasks can be decisive.
False positives are becoming a burden for experts
The surge in interest has an obvious cost. Cryptography specialists are being asked to validate a flood of AI-generated answers, many of them wrong. Elonka Dunin, a retired game developer in the Washington DC area who has written books on cryptography, described the volume of incorrect submissions as overwhelming.
False positives are a predictable byproduct of AI-assisted guessing. Large language models can generate a plausible chain of reasoning even when the underlying solution is wrong, and in codebreaking that can look convincing at first glance. A message can appear nearly decoded while still being keyed incorrectly, mistranscribed or matched to the wrong historical context. For experts, the burden is no longer only solving the puzzle; it is also filtering out the growing number of users who arrive with a polished but invalid answer.
The practical effect is straightforward. More people can now enter historical cryptanalysis, but more expert labor is needed to separate genuine solutions from machine-generated noise.
Why these cases matter beyond hobbyist cryptography
These episodes are less about replacing cryptographers than about changing the access model for technical problem-solving. Large language models can lower the entry barrier to tasks that once required years of specialized training, particularly when the task blends pattern recognition, coding, historical research and patient iteration. In both of these cases, the model did not simply produce text; it acted as a coordinating layer between source material, algorithmic search and domain knowledge.
At the same time, the limitations are obvious. The Enigma breakthrough required confirmation from an independent cryptographer, and the 1809 letter still depended on human judgment to establish that the decoded result made historical sense. Model output alone is not enough.
For now, the most important unanswered issue is not whether AI can assist with codebreaking. It already can. The sharper question is how many incorrect solutions experts will have to review before the new wave of amateur cryptanalysis becomes sustainable.