AI-Assisted Cryptanalysts Report Two Long-Unsolved Enigma Messages Decoded
Two researchers say they have recovered the contents of previously unsolved Enigma messages with help from large language models. TechCrunch describes one attempt using OpenAI’s Astra and another using Anthropic’s Claude Opus 5. The cases are not identical demonstrations of autonomous problem solving: one began with a broad request to locate and decode a message, while the other involved substantially more guidance from a human cryptanalyst.
The reported outcomes are striking partly because these are archival leftovers, not the ordinary Enigma traffic deciphered during World War II. Some messages remain unresolved because their original encoders made mistakes or later records were transcribed incorrectly. Recovering the text can require more than trying combinations on a machine; it can require research into the historical context that narrows the possibilities.
An open-ended task for Astra
Developer Carter Leffen asked Astra to search a database of unsolved Enigma messages and decode one, according to TechCrunch. The model chose a message, researched archival context, assembled clues, built a simulator of the cipher machine and produced plaintext for a message that had resisted researchers since 2005. Leffen also used Astra to construct an interactive explanation of the process. The report presents this as a case in which the model performed both research and technical work, rather than merely completing a calculation supplied by a human.
Frode Weierud, a retired electrical engineer who maintains a collection of cryptology resources and message records, checked Leffen’s solution last week. His validation is an important point of external review in the source’s account. It supports that this particular solution was not simply an unverified model output. It does not, by itself, show that the same approach would solve every remaining message or that all of the model’s intermediate research was reliable.
The model’s logs also raised an unresolved question. They referred to messages in a private collection that Weierud does not host. He was uncertain whether Astra had reached those materials, found them elsewhere online, or encountered related records in public German archives. TechCrunch does not establish which explanation is correct. That uncertainty matters because it limits how precisely readers can describe what information the model actually used.
Claude solved a different message with more direction
The second case came from Jack Willis, a cybersecurity executive and cryptanalyst who contacted Weierud on September 21. Willis used Claude Opus 5 to break a different unsolved message. Unlike Leffen’s broad instruction to Astra, Willis supplied significantly more guidance, TechCrunch reports. The decisive route involved the known signature of a particular officer, which helped the model arrive at the message’s contents.
That difference in human involvement is central to interpreting the pair of results. Both involve AI assistance and an archival answer, but the source does not describe one standardized benchmark or a head-to-head comparison between the models. The human researcher set the task, selected or supplied clues and assessed the result in different ways. Calling both outcomes simply autonomous would obscure how much expertise Willis contributed in the second attempt.
The story also does not claim that Enigma itself has suddenly become easy to break in every circumstance. The historical machines and surviving records present different obstacles. The article says seven Enigma messages remain unbroken, plus one for which the plaintext is known while the underlying cipher has not been recovered. That count is Weierud’s account of the remaining archive, not a measurement of either model’s general cryptanalytic ability.
What the examples establish—and what they leave open
These cases show that a model can help connect archival searching, contextual inference, simulation and candidate decoding in a real research workflow. They are especially informative because the output can be checked against a concrete historical message. In the Astra case, Weierud’s review supplied that check. In the Claude case, TechCrunch reports a separate solution reached with substantial researcher direction.
The account leaves other questions unanswered. It does not provide a controlled measure of how often either system fails, how many unsuccessful attempts preceded the reported successes, or exactly what private or public material Astra saw during its research. Those limits do not erase the two reported solutions. They define what can reasonably be inferred from them: two specific archival messages were recovered with model assistance, through different combinations of machine work and human guidance.
