Two people used AI models to break German Army Enigma messages from 1941 that nobody had solved. GPT-6 Astra broke one that had resisted every attempt since 2005, and Claude Opus 5 broke another. Frode Weierud, who publishes the messages on his Crypto Cellar website, documented both breaks.
New to this? Read it in simple words
- Enigma was a machine that the German army used to hide its radio messages in the Second World War.
- Two people used AI models to break two old messages that nobody had solved.
- One message, from 1941, had resisted all attempts since 2005.
- People guided the AI, and an expert checked the answers.
- Enigma
- A German cipher machine from the Second World War that scrambled letters to hide messages.
- Crib
- A guessed piece of the hidden text, used as a starting point to find the key.
- Bombe
- A wartime machine that tested many Enigma settings quickly to find the right one.
Astra and the message MVUEH
Carter Leffen asked OpenAI’s GPT-6 Astra to see whether it could break any of the unbroken messages on the site. The model chose message number 172, known as MVUEH, which had resisted every attempt since 2005.
It guessed that the text repeated a village name, ROSENOW ROSENOW, and used that as a starting clue. It then wrote its own Enigma simulator and a software version of the Bombe, the codebreaking machine used in the war, and searched for the key.
The work took about two days, on September 14 and 15. The message, sent on 10 July 1941, asks for the route of march and says that the sender is in Rosenow.
Weierud wrote that it was immediately clear the key and the text were correct. “What it has achieved in two days would take a human researcher weeks or even months,” he said.
Documented by Crypto Cellar’s Frode Weierud. People guided both attempts.
Claude and the message FMNGI
Jack Willis, a cybersecurity executive, used Anthropic’s Claude Opus 5 on message number 285, known as FMNGI, from 31 July 1941. He emailed Weierud about the break on September 21.
Weierud says this attack relied on strong human guidance. Willis gave Claude a codebreaking toolkit and the historical records, and the model used an officer’s signature, HARTJENSTEIN, as its clue.
The final key search took 13 minutes and 28 seconds on an Apple M2 computer, according to Weierud’s page.
TechCrunch notes that Leffen gave Astra very little instruction, while Willis guided Claude much more closely.
How much did the AI do alone?
Weierud wrote that Astra did the MVUEH break entirely on its own. Leffen’s own write-up is more careful. It calls the work “researcher-led” and says a related message that was already solved helped suggest the clue.
Astra’s logs also cite files from Germany’s federal archive that are not on Crypto Cellar. Weierud says the references are correct, but he is not sure how the model found them.
Weierud found errors in the old copy of the MVUEH message, which may explain why people failed before. He says seven Enigma messages in his collection are still unbroken, plus one puzzle.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2
- 3
- 4
- 5
We read both break reports on Crypto Cellar, Leffen’s own case study, and Bruce Schneier’s blog, then compared TechCrunch’s article. Weierud checked the solutions; we did not. Accounts differ on how much the people helped, so we give both views.