AI Models Crack 2 Unbroken Enigma Messages, Leaving 7 Still Unsolved

Q: How did an AI model secretly access private WWII archives to crack a decades-old Nazi code in just hours?

There is no evidence the AI “secretly accessed” or hacked private WWII archives. What is documented is that OpenAI’s Astra appears to have used publicly available records, notes on Crypto Cellar, and archival references it inferred or surfaced during research, including Bundesarchiv file numbers and mentions of a “private collection.” Frode Weierud, who verified the Enigma solution, said those references were real but not hosted on his site, so the exact source path remains unclear. What is clear is how the model solved the code. Carter Leffen tasked Astra with examining unsolved Enigma messages. The model selected the 10 July 1941 message MVUEH, linked it to a related solved message, SIPVX, and focused on the repeated place name “ROSENOW ROSENOW” as a likely crib. It then built its own Enigma simulator and Bombe-style software in Python and C++, searched roughly 14.8 million candidate keys, and recovered the machine settings and plaintext in about 10 hours of model time. The break was plausible because the message had hidden clues that earlier researchers missed: transcription errors in the archived ciphertext, a rare left-wheel turnover at the 72nd letter, and a key setup different from other traffic that day. Weierud said Astra behaved like a skilled cryptanalyst and archive researcher, doing in days what might take humans weeks or months. So the breakthrough seems to have come from powerful automated research and cryptanalysis, not proven unauthorized archive access.

1 articles · Updated · TechCrunch · Sep 25

Two cryptanalysts used OpenAI’s Astra and Anthropic’s Claude Opus 5 to decode two previously unbroken Enigma messages, including one that had stumped researchers since 2005.
Astra solved one message after searching Enigma archives, building its own simulator and using contextual clues; Crypto Cellar curator Frode Weierud validated the result last week and said he was in “awe.”
Weierud said Astra’s logs referenced archival material beyond his database, raising questions about whether the model drew on privately shared records or Germany’s public archives.
Claude Opus 5 broke a separate message on Sept. 21 after cryptanalyst Jack Willis gave it more guidance, including the known signature pattern of a specific officer’s name.
Just seven Enigma messages remain unbroken, plus one whose plaintext is known but whose exact encoding has not yet been cracked.