Two researchers used OpenAI’s Astra models and Anthropic’s Claude Opus 5 to decode two messages encrypted by Enigma machines, among seven known unsolved messages that remain. The case demonstrates the models’ ability to combine archival research, simulation, and reasoning, but leaves open questions about the data sources they accessed.
OpenAI’s Astra models and Anthropic’s Claude Opus 5 successfully decoded two historical messages encrypted by Enigma machines after they had remained unsolved for many years. The result represents a practical application of large language models to a task combining cryptanalysis, archival research, and reconstruction of the encryption circumstances.
How were the two messages solved?
Developer Carter Leffen asked the Astra model to search a database of unsolved Enigma messages and attempt to decode one of them. The model conducted archival research, extracted contextual clues, and built an Enigma machine simulator before recovering the original text of a message that had puzzled researchers since 2005. Leffen also used the model to create an interactive website explaining the problem.
Frode Weierud, a retired electrical engineer who runs the Crypto Cellar website and maintains resources, records, and a database of Enigma messages, reviewed and authenticated the solution. He said that what Astra accomplished in two days would have taken a human researcher weeks or months, according to the source’s account of his comments.
As for the second message, it was decoded by Jack Willis, a cybersecurity executive, using Claude Opus 5. Willis contacted Weierud on September 21 and gave the model more detailed instructions, including the name known from the signature of a particular officer, which helped it reach the original text.
Why does this result matter?
The significance of the case lies not merely in a model outperforming a historical cipher machine, but in the nature of the workflow Astra followed: searching for sources, understanding the context, building a simulation tool, and then testing a hypothesis to decode the message. This shows how agentic models can handle problems that cannot be solved with a memorized answer but instead require a sequence of steps and tools.
At the same time, the example alone does not prove that the model carried out independent reasoning or that the solution was free from the risks of accessing external information. Astra’s records include references to messages preserved in a private collection not hosted on Weierud’s website. He was unable to determine whether the model had accessed them, whether the information had been published elsewhere, or whether it appeared in the German government’s public archives. The verifiability of the model’s sources therefore remains an essential part of evaluating the result.
How many Enigma messages remain?
According to Weierud, seven Enigma messages remain unsolved, in addition to one message whose original text is known to researchers but whose cipher they have not deciphered. This list may become shorter if models manage to repeat this type of work, but the source does not yet provide evidence of their success with the remaining messages.