OpenAI's GPT-6 Astra Cracked a Nazi Enigma Message in 10 Hours: Is Bitcoin Next?
A German Army Enigma message from July 1941 has finally given up its contents. Carter Leffen broke it this week using OpenAI’s GPT-6 Astra, and the archive that had logged the intercept as unsolved now records a solution.
The recovered text is almost mundane. Reaching it took roughly 10 hours of model time, one borrowed phrase, and 14.8 million key checks.
Inside the GPT-6 Astra Enigma Break
The intercept carries the indicator MVUEH and dates to July 10, 1941. It came from German Army traffic during the invasion of the Soviet Union. The run recovered the full machine setup, including rotor order II-V-III and 10 plugboard pairs.
A crib made the attack practical. Leffen took the repeated place name Rosenow from a related message that was already solved, which cut the search space to a workable size. GPT-6 Astra then searched archives, built an Enigma simulator, wrote the cryptanalysis code, and tested competing keys in parallel.
“Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio,” Carter Leffen, site
Two independent implementations reproduced the plaintext. Frode Weierud’s CryptoCellar archive has since logged the GPT-6 Astra Enigma break and credited Leffen by name.
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
Where the Human Still Did the Work
The cipher itself is not the story. Allied codebreakers beat Enigma in the 1940s, and any modern machine can clear a three-rotor key space once a crib narrows it.
Nor was the break autonomous. Leffen set the goal, supplied the crib, and pushed the investigation forward, while the agents handled evidence, search, and review. The model carried out the execution rather than the judgment.
That division of labor keeps surfacing. Anthropic said in July that its model found weaknesses in encryption systems that researchers had not documented.
However, none of this touches modern cryptography. Enigma is a 1930s rotor machine, and the math protecting Bitcoin wallets sits in a far harder class. The quantum threat to Bitcoin remains a live question there.
Leffen offered his own measure of the effort. He said the website explaining the GPT-6 Astra Enigma solution took 99 times more work than the codebreaking itself.