A German Army Enigma message from July 1941 that had remained unresolved for decades has reportedly been deciphered with the help of OpenAI’s GPT-6 Astra. The effort, led by Carter Leffen, involved historical research, cryptanalysis, software development and millions of key checks.
The breakthrough offers a striking demonstration of what modern AI agents can do when they combine browsing, coding, reasoning and iterative experimentation. But it does not mean that modern cryptographic systems such as Bitcoin have suddenly become vulnerable.
The decrypted message, identified as MVUEH, was transmitted on July 10, 1941. The recovered plaintext is relatively ordinary: the sender requested instructions for a route of march, stated that he was in Rosenow, and asked for an immediate radio response.
GPT-6 Astra and the 1941 Enigma Challenge
The message was part of a collection of German Army Enigma traffic preserved by Frode Weierud’s CryptoCellar archive. CryptoCellar’s July 1941 records identify message No. 172 by the indicator MVUEH and show that it did not decrypt using the ordinary key for that day’s traffic.
According to the published account of the break, Leffen instructed GPT-6 Astra to investigate unresolved messages in the archive. The AI examined related historical material and identified MVUEH as a promising target.
The task required considerably more than simply running a brute-force search.
GPT-6 Astra reportedly:
- searched historical archives;
- compared uncertain ciphertext readings;
- identified contextual clues;
- developed an Enigma simulator;
- wrote cryptanalysis software;
- tested different approaches;
- performed parallel key searches; and
- independently checked the resulting plaintext.
The reported model runtime was approximately 10 hours.
Two days ago, GPT-6 Astra broke a yet unsolved German Army Enigma message from 1941. Amazingly Astra was able to autonomously:
— Carter Leffen (@carterleffen) September 17, 2026
– Search historical archives
– Compare uncertain letters
– Find contextual clues
– Build an Enigma simulator
– Write cryptanalysis code
– Run parallel… pic.twitter.com/m5usF56bsH
The Rosenow Clue That Helped Unlock the Cipher
One of the most important pieces of the investigation was the repeated place name Rosenow.
A related message from the same period had already been solved and contained the place name. That gave the AI a potential crib—a probable piece of plaintext that could be used to constrain the Enigma search.
This distinction matters.
The achievement was not simply asking an AI to “break Enigma” and waiting for an answer. The system had to investigate the available evidence, determine which information could constrain the search, construct the necessary tools and then test the resulting hypotheses.
The original Enigma system also had an important property: it could not encrypt a letter as itself. That characteristic can be used in cryptanalytic attacks to eliminate impossible crib placements and reduce the number of candidate configurations.
The combination of historical context, the Rosenow crib and machine constraints transformed an enormous theoretical search into a practical computational problem.
What Was Recovered From the MVUEH Message?
The published reconstruction gives the literal machine output as:
BTTE UM ANGABE DES MARSQWEGES X BEFINDE MIQ IN X ROSENOW ROSENOW X SOFORT FUNKANTWORT X WASCHBBSCH
The archive project explains that the unusual spelling is part of the machine output rather than something that should simply be corrected by an editor. The approximate English meaning is:
“Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio.”
The final portions of the German text contain apparent spelling irregularities, including “BTTE” and “WASCHBBSCH.” The project says these details were preserved rather than silently corrected because they form part of the cryptanalytic evidence.
The recovered solution also included the Enigma machine configuration. CryptoCellar’s records identify the relevant wheel order and plugboard settings for message No. 172.
14.8 Million Key Checks
The scale of the computational work is another notable part of the story.
An independent account of the reconstruction reports that approximately 14.83 million physical key configurations were checked against the message header, with further filtering used to identify configurations compatible with the recovered plaintext and other evidence.
That figure should not be confused with simply testing every possible Enigma configuration from scratch.
The historical clue, ciphertext structure, known header information and other constraints dramatically narrowed the practical search. The AI’s contribution was therefore a combination of cryptanalysis, software engineering and evidence management, rather than raw computational power alone.
Where Human Judgment Still Mattered
The story is also a useful example of the current relationship between humans and AI agents.
Although GPT-6 Astra reportedly performed much of the technical execution, a human initiated the investigation and defined the overall objective. Leffen also supplied direction and reviewed the emerging results.
The CryptoCellar project says that the investigation involved surviving copies of the message, uncertain handwritten characters and independent checks of the recovered configuration.
That makes the achievement different from a completely unsupervised discovery.
A better description is that an AI agent handled a large portion of the research and execution loop, while humans established the goal and evaluated the evidence.
This distinction is increasingly important as AI systems become capable of operating computers, writing software and conducting long-running research tasks.
Why GPT-6 Astra Is Relevant to the Story
OpenAI describes GPT-6 Astra as its most capable model, with capabilities spanning computer use, browsing, software engineering, science and professional work. The model also supports tools including web search, code execution, computer use and hosted shell environments.
Those capabilities are particularly relevant to a task such as Enigma cryptanalysis.
A conventional chatbot could explain how an Enigma machine works. An agentic system can go considerably further: it can locate source material, write a simulator, execute experiments, inspect results and revise its strategy.
That ability to move between research and action is arguably more significant than the individual cipher itself.
Does This Mean AI Can Break Bitcoin?
This is where the story requires some caution.
The successful recovery of a 1941 Enigma message does not demonstrate that AI can break the cryptography protecting Bitcoin.
Enigma was a mechanical cipher system from the World War II era. Modern cryptocurrency security relies on fundamentally different mathematical constructions and cryptographic assumptions.
Bitcoin’s signatures, for example, depend on elliptic-curve cryptography rather than the rotor-and-plugboard mechanism used by Enigma.
Consequently, solving an old Enigma message does not provide a direct path to recovering private keys from Bitcoin wallets.
The more relevant question is whether increasingly capable AI systems could eventually assist researchers in finding weaknesses in modern cryptographic implementations or algorithms.
That is a much broader cybersecurity question.
OpenAI itself describes GPT-6 Astra as having significantly advanced cybersecurity capabilities and says the model reaches the “Critical” threshold under its Preparedness Framework.
The Real AI Breakthrough May Be the Workflow
The most interesting part of the MVUEH story may therefore not be that an AI decoded an old German military message.
It is the workflow.
GPT-6 Astra reportedly moved through several stages that traditionally required different specialists:
Historical research → evidence gathering → hypothesis formation → software development → cryptanalysis → large-scale testing → verification
That type of end-to-end workflow is exactly where agentic AI systems are beginning to show their potential.
Instead of merely answering questions, an advanced model can increasingly become an active research partner capable of building the tools required to answer those questions.
What the Enigma Break Actually Proves
The MVUEH case provides a compelling demonstration of AI-assisted cryptanalysis, but it should not be interpreted as proof that modern encryption is obsolete.
Instead, it demonstrates three more specific things.
First, advanced AI can combine information from multiple sources to formulate useful cryptanalytic hypotheses.
Second, AI agents can create and execute specialized software as part of a larger investigation.
Third, difficult technical problems can become substantially more accessible when an AI system can repeatedly research, code, test and revise its approach.
The Enigma machine was defeated long before today’s AI existed. What is new is the possibility of giving a modern AI agent access to the entire problem-solving pipeline.
The Bigger Question for AI
The GPT-6 Astra Enigma episode may ultimately be remembered less for the historical plaintext and more for what it says about the evolution of AI agents.
A system that can research an obscure historical problem, construct its own technical tools, perform millions of computational checks and independently reproduce a result represents a different category of AI capability from a traditional question-and-answer chatbot.
At the same time, the episode shows why verification remains essential.
Historical documents can contain transcription errors. Cryptographic searches can produce false positives. A plausible plaintext is not automatically proof of a correct key.
In this case, the published project describes independent implementations and additional checks intended to validate the recovered solution.
The Enigma message may be 85 years old, but the broader lesson is current: AI agents are increasingly capable of turning complex research questions into executable technical investigations.
And while Bitcoin is not suddenly at risk because an Enigma message was solved, the rapid improvement of AI-assisted cybersecurity means the relationship between artificial intelligence and cryptography deserves much closer attention.
Discover more from AiTechtonic - AI & Informative News
Subscribe to get the latest posts sent to your email.