A German Army Enigma message sent on 10 July 1941 may no longer be among the unresolved traffic documented by Crypto Cellar Research. In an account updated on 19 September 2026, cryptologist Frode Weierud says Carter Leffer contacted him on 15 September with a proposed break of message MVUEH, designated Nr. 172. Weierud reports that the message had resisted previous attempts since 2005 and that the proposed solution immediately produced a coherent key and plaintext. That is a strong expert assessment, but it remains an account published by a participant rather than an independently reproduced result.
The historical record gives the message a precise setting. Weierud says radio station 2ny transmitted it and that the SS-Totenkopf Quartiermeister station, Ib, received it at 17:30 and entered it as Nr. 172 in its July incoming-message log. The ciphertext contains 82 letters. According to the account, its recovered key used wheel order 253, unlike wheel order 512 associated with other traffic from that day. This matters because an analyst who assumed that MVUEH shared the ordinary daily key would have been searching the wrong configuration space.
The strongest part of the proposed solution is not simply that the ciphertext can be transformed into readable German. The recovered plaintext is reported to be almost identical to that of Nr. 173, SIPVX, another message from the same date that Alex Shovkoplyas broke in 2017. MVUEH is 82 letters long and SIPVX is 94. Weierud attributes the difference to an enciphering error that turned Bitte into Btte, together with a repeated sender signature in SIPVX. A solution that explains both the shared wording and these specific length differences is more persuasive than a decryption that merely happens to resemble plausible language, although the underlying images and full verification record were not included in the supplied material.
The account also identifies obstacles that could explain why the message remained difficult. Weierud says the transcription copied from the original message form contained several errors. He also reports that the Enigma’s left-hand wheel turned over at the 72nd letter, an uncommon event that can complicate a crib-based attack. Those details offer a concrete historical explanation for earlier failures: the surviving data were imperfect, while the machine’s stepping behavior departed from the simpler assumptions an analyst might make. They do not, however, prove that these were the only reasons the message survived unsolved.
The central claim about GPT-6 Astra requires especially careful wording. Weierud says Leffer asked the system to examine the unbroken messages on the Crypto Cellar Research website, after which it selected MVUEH as promising, suspected a relationship with SIPVX, and developed Python and C++ tools including an Enigma simulator and Bombe. The account says Astra then used the repeated place name ROSENOW ROSENOW as a crib and searched until it found the reported key and plaintext. This describes a potentially capable research workflow, but not a model decrypting the message inside its neural network. The substantive computation was performed by generated programs running on conventional hardware.
Paragraph 5 explains that GPT-6 Astra, an OpenAI system, helped direct the reported Enigma research workflow. Credit: ja.wikipedia.org.
That distinction answers the practical question of what credit an AI system deserves. Astra may have contributed by locating relevant archive references, forming the cross-message hypothesis, choosing a crib, and directing a search. The supplied account does not quantify how much of that reasoning was original, how much came from publicly available material, or how many human corrections occurred between attempts. Nor does it provide the complete logs or software needed to reproduce every step. The evidence therefore supports calling Astra a substantial assistant in the reported break, not proving that it independently solved the cipher without meaningful human guidance.
Archive research adds another layer of uncertainty. Weierud says Astra located references to German Bundesarchiv radio-message volumes RS 3–3/20a and RS 3–3/63b, references he confirms are valid even though they are not hosted on the research site. He also says the original image for received message No. 172 had not been located in the available environment and that the meaning of Astra’s reference to a private collection remained unclear. Those unresolved points matter because a verified image of the source form would allow researchers to check the disputed transcription directly. Until that provenance is documented, the historical identification and the exact ciphertext transcription should be treated as reported findings rather than settled archival facts.
For readers outside cryptography, the episode is less a demonstration of mysterious machine intuition than an example of agentic research: a language model can move between documents, propose connections, write specialized software, run searches, and revise its approach. Its apparent strength here came from combining several ordinary capabilities around a tightly constrained historical problem. The result is still impressive if the account is accurate, but it should not be generalized into evidence that current AI systems can routinely solve unfamiliar ciphers, validate archival documents, or conduct reliable research without expert oversight.
The appropriate conclusion is deliberately narrower than the original headline. The supplied primary account presents a coherent proposed solution, a plausible explanation for past difficulty, and an expert’s judgment that the key and plaintext are correct. It does not yet supply independent reproduction, the original No. 172 image, complete execution logs, or enough detail to separate the model’s contributions from human direction and conventional computation. Those missing pieces are the next tests of the claim—not reasons to dismiss it, but reasons to describe it precisely while verification continues.