The Enigma decrypt was accepted on a header held out of the key search
A Bloomberg developer's agent run recovered an 82-letter German message from 1941 in about ten hours. What makes it checkable is an acceptance procedure that enumerated every doubtful letter before the search started.
The Engineer · Build desk
What happened
- Carter Leffen, a product development coach at Bloomberg LP in New York, says the GPT-6 Astra Extra High variant worked about ten hours on MVUEH, an 82-letter German Army message sent on July 10, 1941.
- The search leaned on a 14-letter crib, ROSENOWROSENOW, taken from a message solved earlier the same day; settings that could not produce those letters were eliminated.
- Several handwritten characters in the surviving copy were unclear even against a published transcription, so the investigation identified 12 positions where more than one letter was possible.
- Acceptance required the recovered key to decode all 82 reconstructed letters and to match information recorded separately from the message body, after which the code and search materials were published.
- SWARM published its own review on September 17 using a separate simulator, which reproduced the expected message and header result without repeating the full key search.
Why it matters
- What it limits constraint Coherent German is the cheapest thing a language model produces, so the 68 letters outside the crib had to carry the accept-or-reject decision.
- Who has to decide decision Enumerating the doubtful letters up front takes away the option of re-reading a smudged character once one reading yields a better sentence, and it gives a reviewer a fixed candidate set to check the answer against.
- What doesn't add up contradiction the-decoder reports that the separately archived header confirmed the settings; theneuron.ai's account has the same header leaving 923 candidates standing, so a reader weighing the evidence gets two different strengths for the same check.
- What it sets up precedent Shipping the code, the search data and a simulator means reviewers can replay the calculation instead of taking the transcript on trust, and it sets that as the expected form for AI-assisted historical claims.
Next story · Build · 6h ago
Gemini 4 Argon falls from 97.5% to 49.8% on ARC-AGI-3 once vendor features come offConfirmed2 publishers
German operators wrote the machine setup data into the message header, and the original key search never used it, according to theneuron.ai's account of the investigation [6]. That left the header as held-out evidence. Tested afterwards, 14,829,646 possible Enigma setting combinations yielded only 923 that produced the expected header result [7]. One candidate in roughly 16,000 survived [26], and the final key still had to be identified some other way [7].
Those twelve doubtful positions allow 13,824 versions of the ciphertext [16]. That number factors as 2^9 times 3^3, which is exactly twelve prime factors, so nine of the positions had two candidate letters and three had three [27]. The list was written down before the answer existed [17].
Brute force was hopeless against the roughly 159 quintillion possible daily settings the machine offered [21]. Known keys from the same day did not fit, and automated searches returned nothing useful, the-decoder reports [22]. The crib worked because of a design flaw: Enigma never encrypts a letter as itself, so an alignment that puts a crib letter over the same ciphertext letter can be discarded at once [9]. At one alignment everything held, and the 68 letters outside the crib came out as German, including "Sofort Funkantwort" and "Angabe des Marschweges" [10]. The project builds on earlier cryptanalytic work by Frode Weierud, Geoff Sullivan and Olaf Ostwald [12].
Then the checks. Separate programs reproduced the message and the header result, and another verification system matched machine behaviour across 42 sample cases [18]. Re-encryption turned the recovered plaintext back into the original ciphertext. That check showed the calculation ran in both directions, and it did not establish the historical reading [20]. A second search over the published transcription found the same key, and because it reused a phrase discovered during the first recovery, the project logs it as confirmation of the first result [19].
The two published accounts differ on how much the header proved. the-decoder writes that a header preserved separately in the archive confirmed the settings [11]. theneuron.ai describes it as a check applied after the search, one that eliminated most alternatives and left the key itself unidentified [7].
Leffen's own argument for authenticity is the typing. The plaintext reads BTTE where BITTE belongs and WASCHBBSCH where the sender's name probably was, and he takes the errors as evidence the decryption is real, since a fabricated result would more likely be error-free [23]. The signature remains uncertain [3]. On effort, he said he put "99 times more effort into building the website that describes the problem and the solution than into actually cracking the code" [24].
This bar was reachable because the archive supplied a channel of evidence the search could not consult, and because the ambiguity in the source document could be enumerated and frozen before the model began producing plausible sentences. Most agent workflows have neither. The code, the search data and a working 3D Enigma simulator can be downloaded from the project page. The packages verify the calculations and replicate the key search sequence; what they leave open is the message's historical identity and whether the solution is the only possible one [25].
What to watch
- Whether anyone repeats the full key search from the published transcription without reusing a phrase found during the first recovery.
- Whether reviewers working through the published packages find a second setting that also yields coherent German across all 82 letters.
- Whether the uncertain WASCHBBSCH signature is resolved against records kept separately from the message body.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence62
- Adoption22
- Hype gap+26
- Incentives52
- Confidence60
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
A September 14 to 15 investigation used GPT-6 Astra and several specialist AI agents to recover MVUEH, an 82-letter German Army message from July 10, 1941; the agents reviewed old documents, built search programs, tested possible answers and challenged results.
- [2]
According to a detailed case study, Carter Leffen, a product development coach at Bloomberg LP in New York, decrypted the message using OpenAI's GPT-6 Astra; the "GPT-6 Astra Extra High" variant worked about ten hours on the problem, Leffen says.
- [3]
MVUEH was German Army message No. 172; its recovered text roughly translates to "Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio." The apparent signature, "Waschbusch," remains uncertain.
- [4]
A previously solved message contained the repeated place name ROSENOWROSENOW; researchers suspected the same 14 letters appeared in MVUEH and used it as a crib, eliminating settings that could not produce ROSENOWROSENOW.
- [5]
The remaining 68 letters still had to produce coherent text, so a setting could not be accepted just because it matched the guessed place name.
- [6]
The message header contained data operators used to set up the Enigma machine before encrypting the main text; since the original search did not use the header, researchers could use it afterward as an independent check on the recovered settings.
- [7]
Across 14,829,646 possible Enigma setting combinations tested against the header, only 923 produced the expected result; the header did not identify the final key on its own, but it eliminated most alternatives.
- [8]
External review has begun, with SWARM publishing its own review on September 17 using a separate simulator; it reproduced the expected message and header result, although its authors did not repeat the full search, so the review confirms only part of the investigation.
- [9]
The machine never encrypts a letter as itself, so if the plaintext contains an R the ciphertext at that position can be anything except R; that alone let them rule out many positions right away.
- [10]
At one position everything lined up, the team calculated the matching machine settings, and the remaining 68 characters produced coherent German text including "Sofort Funkantwort" and "Angabe des Marschweges".
- [11]
A message header preserved separately in the archive confirmed the settings.
- [12]
The work builds on cryptanalytic contributions by Frode Weierud, Geoff Sullivan and Olaf Ostwald.
- [13]
Whether the decryption holds up to independent review will depend on experts evaluating the published materials.
- [14]
Enigma was broken decades ago by Polish and Allied cryptanalysts; this project focused on one message that had remained unresolved because researchers did not know the exact machine settings used to encrypt it.
- [15]
Getting a coherent German translation was not enough to call the message solved: the recovered Enigma key also had to decode all 82 reconstructed letters and match information recorded separately from the main message. Once those checks lined up, the researcher published the code and search materials.
- [16]
Several handwritten characters were hard to read even with a published transcription available, so the investigation identified 12 positions where more than one letter was possible; combined, those uncertainties created 13,824 possible versions of the source text.
- [17]
Researchers recorded those possibilities before finding the answer, which reduced the chance that they would reinterpret an unclear letter simply because another choice produced a more convincing sentence.
- [18]
Separate programs produced the same message and header result, while another verification system matched the machine behavior tested in 42 sample cases.
- [19]
Using the published transcription, a second search also found the same key, although it reused a phrase discovered during the first recovery; the project treats the result as confirmation of the earlier finding rather than an independent second discovery.
- [20]
Re-encryption tested whether the recovered settings could turn the decrypted text back into the original encrypted text; passing showed the calculation worked in both directions, but it did not prove the researchers had found the correct historical interpretation.
- [21]
The Enigma machine offered roughly 159 quintillion possible daily settings, so brute-forcing all of them was never an option.
- [22]
Several approaches failed early on: known keys from the same day did not fit, and automated searches returned nothing useful.
- [23]
The decrypted text contains typos that likely came from the radio operator, such as "BTTE" instead of "BITTE" and "WASCHBBSCH" instead of what was probably "Waschbusch"; Leffen sees these mistakes as evidence the decryption is authentic, because a fabricated result would more likely be error-free.
- [24]
Leffen says he put "99 times more effort into building the website that describes the problem and the solution than into actually cracking the code."
- [25]
All code, search data and a working 3D Enigma simulator are available for download on the project page; the provided packages verify the calculations and replicate the key search sequence, but they do not prove the message's historical identity or that the solution is the only possible one.
- [26]
About one candidate in 16,000 survived the header test.
- [27]
13,824 equals 2^9 x 3^3, which is exactly twelve prime factors, so the 12 ambiguous positions must be nine positions with two candidate letters and three with three candidate letters.
Sources
2 independent publishers whose own reporting we read for this story.
- the-decoder.comOpenAI's GPT-6 Astra decrypts a Nazi radio message in ten hours that went unsolved for 83 years
1 article · September 17, 2026
- theneuron.aiHow AI Cracked an 85-Year-Old WWII Enigma Message
1 article · September 18, 2026
Topics and entities
Topics
- AI agent workflows
- WWII signals intelligence
- Research Reproducibility
- Enigma cryptanalysis
Entities
- OpenAI
- MVUEH (message No. 172)
- GPT-6 Astra