Astra went beyond simply testing possible keys. According to the account of the work, the model researched historical records, gathered contextual clues and created its own Enigma simulator before recovering the plaintext.
Leffen then used Astra to build an interactive website explaining the process and the resulting solution.
Frode Weirerud, a retired electrical engineer who maintains the Crypto Cellar archive of Enigma material, reviewed the result and confirmed the solution last week. He said the work left him in “awe.”
The second break came from cryptanalyst Jack Willis, who contacted Weirerud on September 21 after using Anthropic’s Claude Opus 5 to solve another outstanding message.
That case involved more direct human assistance. Willis supplied Claude with the known signature of a particular officer, giving the model an important clue that helped it work backward toward the plaintext.
The two cases illustrate different levels of AI involvement. Astra handled much of the research and tooling itself, while Claude’s successful attempt depended more heavily on expert guidance.
The speed of Astra’s work also stood out. Weirerud said he had previously spent several weeks examining some of the same Bundesarchiv material that the model worked through in two days.
There is still some uncertainty around the sources Astra used. Weirerud said the model’s logs referred to archived messages from a “private collection” that is not hosted by Crypto Cellar. He could not determine exactly where that material came from, suggesting it may have been available elsewhere online or through Germany’s public Bundesarchiv records.
Enigma is closely associated with the Allied codebreaking effort during World War II, including the work of Alan Turing and others at Bletchley Park. The Bombe machine was designed to search through possible Enigma settings at scale.
The newly solved messages present a different kind of problem. Many of the small number of Enigma communications that survived without solutions did so because of transcription mistakes or errors made by the original operators, rather than because the underlying encryption itself remained secure.
That makes historical context and archival research especially important. In Astra’s case, the model was not simply searching a clean cryptographic key space. It also had to identify a useful target, investigate supporting material, build software and test possible explanations.
The Claude result was more collaborative. Willis supplied a key contextual clue, and the model used it as part of the decryption process.
Those differences matter when judging how independently the systems performed. Neither case demonstrates that current AI models can solve arbitrary cryptographic problems without assistance. Instead, they show how models can combine coding, research and iterative reasoning when the problem has a defined answer that can be independently checked.
Researchers have now confirmed both solutions, narrowing the remaining set of unresolved Enigma material. Seven messages are still unbroken, while one additional message has known plaintext but an unrecovered code.
The breakthroughs also provide a clear example of AI being used as a research tool rather than only as a conversational assistant. In these cases, the models helped move between archives, historical context, software construction and hypothesis testing — tasks that traditionally would have required substantially more manual work from a human cryptanalyst.
This analysis is based on reporting from NotebookCheck.
Image courtesy of Cryptocellar.
This article was generated with AI assistance and reviewed for accuracy and quality.
