Two cryptanalysts have used frontier language models from OpenAI and Anthropic to decode two long-unsolved messages encrypted by Nazi Germany's Enigma machines during World War II, including one that had resisted human researchers since 2005. The breaks leave just seven unsolved Enigma messages in the known archival record, plus one where the plaintext is known but the code itself remains unbroken.
Developer Carter Leffen prompted OpenAI's newest model, Astra, to search a database of Enigma messages, pick an unbroken one, and decode it. Astra did archival research, gathered context clues, built its own simulator of the Enigma machine, and recovered the plaintext of the 2005 message. Leffen then used Astra to build an interactive website walking through the solution.
The validation came from Frode Weirerud, a retired electrical engineer who maintains Crypto Cellar, a long-running repository of Enigma records and unsolved messages. Weirerud confirmed Leffen's solution last week and said it left him in "awe."
Key facts
- 01Developer Carter Leffen used OpenAI's Astra to decode an Enigma message that had gone unsolved since 2005.
- 02Cryptanalyst Jack Willis used Anthropic's Claude Opus 5 to break a second unsolved wartime message on September 21.
- 03Only seven unbroken Enigma messages remain, plus one where the plaintext is known but the code is not.
- 04Astra built its own Enigma simulator, ran archival research, and reached a solution in two days.
On September 21, a second cryptanalyst, Jack Willis, contacted Weirerud with a different break. Willis had used Anthropic's Claude Opus 5 to crack another unsolved message, providing more hands-on guidance than Leffen did. Claude Opus 5 exploited the known signature of a particular officer's name to work back to the plaintext.
Weirerud noted that Astra's own logs referenced archived messages held in a "private collection" not hosted on Crypto Cellar. He isn't certain how the model accessed them, speculating that another researcher may have posted them online, or that Astra pulled from the German government's public Bundesarchiv holdings.
The scale of the compression is the striking part. Weirerud, who has spent decades on this material, said he personally spent several weeks working through the same Bundesarchiv files that Astra parsed in two days.
The original Enigma break, led by Alan Turing at Bletchley Park, required the construction of the Bombe — an early electromechanical computer purpose-built to search Enigma's key space. The United Kingdom's ability to read German military traffic during the war is generally regarded as the founding case for machine-assisted cryptanalysis. The messages that survived unbroken into the present day did so largely because of mistranscription or encoding errors by the original operators, not because the underlying cipher held.
What Astra and Claude Opus 5 did is different in kind from what the Bombe did. The Bombe brute-forced a known cipher structure. The models handled the full workflow a human cryptanalyst would: locating candidate messages in archives, reasoning about likely context and senders, building the tooling to test hypotheses, and iterating to a solution. That is closer to the job of a research analyst than to the job of a codebreaker.
Two caveats are worth flagging. Willis's break required significant human scaffolding — he supplied the officer's-name signature that unlocked the message, so Claude Opus 5's role was closer to assisted decryption than independent discovery. And the provenance of some archival material Astra referenced is not fully accounted for, which matters for reproducibility even if it does not affect the validity of the plaintext itself.
The broader signal is what these models can now do when pointed at narrow, verifiable problems with a clear success condition. Cryptanalysis of historical ciphers is a good testbed: the answer is checkable, the corpus is bounded, and the skill ceiling is high. Expect more of the remaining seven unbroken Enigma messages to fall, and expect the same pattern — LLM as tireless research analyst, human as validator — to migrate into live scientific problems where the checkable-answer property still holds.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
