“Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio.”
That’s the message sent on July 10, 1941 by a German transport operator during the Operation Barbarossa invasion of Russia that was encrypted on a Nazi Enigma coding machine. Until now, the message sent to SS Death’s Head division headquarters asking to be pointed in the direction of Moscow was a secret. The encryption was broken by AI.
Computer pioneer Alan Turing, best known for his test to distinguish artificial and human intelligence, headed a band of cryptanalysts at Bletchley Park in the UK during World War II. They built the Bombe machine that broke the Enigma encryption scheme. The decrypted intelligence, codenamed Ultra, allowed the Allies to read German coded messages, a breakthrough estimated to have saved millions of lives and shortened the war.
There remained, however, a handful of Enigma messages that were never decoded. That got Carter Leffen, a developer, thinking. Could AI break the Enigma code? Leffen then queried OpenAI’s GPT-6 Astra model to see if it could break any of the unbroken Enigma messages published on the Crypto Cellar Research website.
A clue, called a crib in coding circles, quickly surfaced from a message already decoded that used the repeated place name Rosenow, a city in northeastern Germany. That clue made the search parameters to decipher the “MVUEH” message reasonable in scope. What happened next left Frode Weierud, who maintains the Crypto Cellar Research site, in “awe.”
GPT-6 Astra did its own archival research, found context clues, built a simulator of the Enigma machine and ultimately decoded the message that had baffled human researchers since it became publicly accessible in 2005.
“GPT-6 Astra is behaving like a very professional cryptanalyst and archive researcher,” commented Weierud in a post. “What it achieved in two days would take a human researcher weeks or even months.”
This is not the first time AI has helped unravel a Nazi mystery. German historian Jurgen Matthaus used AI facial recognition and image analysis in 2025 to identify an anonymous Nazi executioner in an infamous Holocaust execution photo. The executioner was identified as Jakobus Onnen. AI also clarified the incident took place in Berdychiv, Ukraine. Conversely, AI has been used to generate fake images related to the Holocaust.
AI also holds promise in generating an authoritative database of stolen Nazi artwork. One such effort is the AI Provenance Assistant being developed at Santa Clara University in California. It’s estimated around 600,000 art works were stolen by the Nazis although a central data repository is lacking.
Meanwhile, GPT-6 Astra’s analysis of the MVUEH message introduced some mystery of its own. Astra’s model logs include discussion of archived messages in a “private collection.” It is unclear whether GPT-6 Astra accessed them or not and whether that data comes from another private researcher or German government archives.
In late September, another cryptanalyst named Jack Willis broke another Enigma message using Anthropic’s Claude Opus 5 model, according to Weierud.
Weierud says there are seven unbroken Enigma messages that remain unbroken.
That number may not last.
