Bloomberg Developer Says OpenAI's GPT-6 Astra Cracked an 83-Year-Old Nazi Enigma Message in 10 Hours

The claim

A Bloomberg developer says OpenAI's GPT-6 Astra decrypted an Enigma-encoded Wehrmacht radio message from July 10, 1941, that had remained unsolved for 83 years. Carter Leffen, a product development coach at Bloomberg LP in New York, published a detailed case study describing how the "GPT-6 Astra Extra High" variant spent roughly ten hours on the problem before producing coherent German text.

The 82-character message, tagged MVUEH in an archive of intercepted German army radio traffic, was sent by a soldier reporting his location in the town of Rosenow and requesting marching orders and an immediate radio reply.

How the decryption reportedly worked

According to Leffen, GPT-6 Astra and a set of specialized AI agents handled multiple parallel tasks: searching historical archives, comparing uncertain letters in the intercepted text, building a working Enigma simulator, writing cryptanalysis code, testing competing key candidates, and cross-checking results against already-solved messages from the same period.

The Enigma machine had roughly 159 quintillion possible daily settings, ruling out brute force. The breakthrough, Leffen says, came from a separately decrypted message sent the same day that contained the town name "Rosenow" twice. The team guessed the name might also appear in the unsolved message and used it as a search anchor, combined with the known Enigma property that no letter can ever encrypt to itself — a constraint that eliminates many possible letter positions immediately.

At one position, the pieces reportedly aligned. Calculating the corresponding machine settings produced legible German text for the remaining 68 characters, including phrases like "Sofort Funkantwort" (immediate radio reply) and "Angabe des Marschweges" (specify the march route). A message header preserved separately in the archive is said to confirm the derived settings. The decrypted text also contains apparent operator typos — "BTTE" instead of "BITTE" and "WASCHBBSCH" instead of a likely "Waschbusch" — which Leffen cites as evidence the result is authentic rather than fabricated.

What hasn't been verified

This is a claim from an individual developer, not an OpenAI announcement or a peer-reviewed cryptography result. Leffen has published code, search data, and a 3D Enigma simulator for download, which allow others to verify the calculations and replicate the key search. But per the case study itself, these materials confirm the math works — not that the historical identity of the message is correct or that this is the only possible solution. Independent review by cryptography historians, building on prior work by Frode Weierud, Geoff Sullivan, and Olaf Ostwald, has not yet been reported.

Notably, Leffen says the AI-driven codebreaking was not the hardest part of the project: he claims to have spent "99 times more effort" building the website documenting the problem and solution than on the decryption itself.

What this means

This is a capability demonstration, not a verified historical or cryptographic milestone. The interesting part isn't Enigma-breaking — that was solved decades ago and this message succumbed to known weaknesses (the no-self-encryption property, a reused daily key context) rather than a new cryptanalytic method. What's notable is the workflow: a human directing multiple AI agents through archive research, simulator-building, code generation, and parallel hypothesis testing over a sustained ten-hour session with minimal supervision. Whether that holds up depends on independent experts checking the published simulator and archive cross-references — something that hasn't happened yet.