Alan Turing's wartime work on the German Enigma cipher laid the foundation for modern computing, yet a handful of intercepted messages have never been deciphered, often because of transcription errors or missing data. Historians and cryptologists continue to catalogue these gaps, noting that only a few messages still lack a readable plaintext after decades of effort.
Developer Carter Leffen instructed OpenAI's newest model, Astra, to locate an undecoded entry within a public Enigma archive and to translate it. The system conducted its own archival search, extracted contextual clues, constructed a functional Enigma simulator, and ultimately produced the clear text of a message that had eluded experts since 2005. Leffen also used the model to generate an interactive website that documents the entire investigative workflow.
Frode Weierud, maintainer of the Crypto Cellar site, confirmed Leffen's solution and described his reaction as one of "awe," according to his statements on the platform. He noted that Astra's internal logs referenced a "private collection" of archived messages not hosted on Crypto Cellar, leaving open the question of whether the model accessed external repositories or public German archives. Leffen later wrote that reproducing the result would have required weeks or months of manual research, yet the model completed the task in two days.
In a separate effort, cybersecurity executive Jack Willis turned to Anthropic's Claude Opus 5 model, providing detailed guidance that included the known signature of a specific officer. Leveraging that identifier, the model succeeded in breaking another long-standing Enigma cipher, delivering a plaintext that had previously resisted conventional analysis. Willis reported the breakthrough to Weierud, adding to the growing evidence of AI-driven cryptanalytic capability.
Weierud reports that only seven Enigma messages remain completely undeciphered, with an additional transmission whose content is known but whose encoding still defies reconstruction. The recent AI-assisted successes suggest that large language models may soon resolve the remaining gaps, potentially completing the historical record of wartime German communications.