Two wartime messages encrypted with Enigma have been decoded with the help of AI models from OpenAI and Anthropic. For software and business leaders, the story offers a compelling example of AI combining research, code generation and experimentation to solve a problem with a verifiable answer.
Alan Turing’s connection to this achievement is particularly fitting. Alongside his foundational work on machine intelligence, he played a central role in the wartime effort to decipher Enigma traffic.
Decades later, researchers are still working through surviving messages. Some present difficulties that go beyond finding the correct machine settings: handwritten records, transmission mistakes and transcription errors can all complicate the analysis.
Two messages, two approaches
Developer Carter Leffen asked OpenAI’s GPT-6 Astra to investigate the unsolved Enigma messages published by Crypto Cellar Research.
According to cryptology researcher Frode Weierud, Astra selected MVUEH, a German Army message from July 1941 that had resisted attempts to decode it since 2005. It identified a potentially related message, developed simulation and cryptanalysis software, and used a repeated place name as a clue to recover the encryption key and plaintext. Weierud confirmed that the solution was correct.
A second effort used a more guided approach. Jack Willis supplied Anthropic’s Claude Opus 5 with historical material and a cryptanalysis workbench. An officer’s signature provided a likely fragment of plaintext, known as a “crib”, which helped narrow the search.
That investigation recovered the message FMNGI. Weierud’s account reports that the final key search took approximately 13 minutes, following the preliminary research and preparation. That distinction matters: the search time represents one stage of the work, rather than the entire investigation.
A model that can build its own tools
The most useful aspect of these cases is how the work progressed.
The AI systems connected historical evidence with computational methods. They helped formulate hypotheses, prepare tools, run searches and examine results. In Astra’s case, the model also developed the software needed to test its approach.
For engineering teams, this suggests a valuable direction for AI adoption: supporting investigations that span several kinds of work.
Understanding an unfamiliar application, for example, can require reading documentation, tracing dependencies, writing diagnostic scripts and testing assumptions. The Enigma cases illustrate a comparable pattern of investigation, although they do not establish how reliably these models would perform across other tasks.
The opportunity is to shorten the path from an unclear problem to a testable explanation.
Verification gives the result its value
These achievements carry weight because the recovered keys and messages could be checked.
That is a useful principle for any organisation introducing AI into technical work. A convincing explanation becomes more valuable when it comes with evidence: a reproducible test, a working implementation or a result that an expert can independently inspect.
The two investigations also show that human involvement can take different forms. One began with a broad instruction; the other relied on substantial preparation and guidance. Both benefited from specialist assessment.
Teams should therefore design AI workflows around the problem, the available evidence and the consequences of an error. The appropriate level of autonomy will vary.
Research needs a traceable evidence trail
One detail in the Astra investigation remains unresolved. Weierud noted that the model’s logs referred to relevant German archival records that were unavailable on his website. The references were accurate, but he could not establish exactly how the model had found them or which material it had accessed.
For businesses, this raises a practical question: can the team trace the evidence behind an AI-generated conclusion?
Source records, tool logs and reproducible steps help reviewers understand what supports a result and where uncertainty remains. This becomes especially relevant when an investigation involves internal documents or sensitive operational information.
What this means for Control F5 Software
For Control F5 Software, the practical lesson connects directly to clarifying complex systems and turning AI initiatives into useful operational software.
An effective AI workflow needs a defined objective, relevant context, appropriate tools and a clear way to assess the outcome. Those foundations help teams apply AI to real engineering problems while keeping decisions grounded in evidence.
The Enigma breakthroughs demonstrate what becomes possible when research and execution come together. For CTOs, the next step is to identify where that combination can resolve a meaningful bottleneck, then build a workflow whose results the team can verify.
We have helped 20+ companies in industries like Finance, Transportation, Health, Tourism, Events, Education, Sports.
