GPT-6 Astra decrypts 1941 Enigma message in ten hours

In September 2026 Carter Lefen launched GPT-6 Astra Extra High on a German army radio transmission MVUEH from 10 July 1941. The message had been archived in CryptoCellar since 2005 and remained undecrypted. After about ten hours of work Astra found Enigma settings and restored the plain text. The radio request asked for a route from Rosenow and demanded an urgent radio reply. The text of the transmission contained no military secret. What matters is how Astra reached it. The message consists of only 82 letters. For cryptanalysis this is a problem rather than a gift because a short ciphertext lacks statistical material. Enigma offered a huge space of possible settings. Finding rotor order, initial positions, ring settings and Steckerbrett connections makes brute force computationally meaningless. Known Enigma settings for other messages that day did not fit MVUEH. Therefore Astra did not rely on brute force but on context search. An archive radio transmission from the same day SIPVX number 173 was found. Researcher Alex Shovkoplyas decrypted it in 2017. Its plain text contained the sequence ROSENOW repeated twice. From this a hypothesis emerged that messages from the same district and time might contain ROSENOWROSENOW. In Enigma cryptanalysis such a hypothesis is called a crib.

Instead of testing all key combinations the search focused on settings that encrypt the known fragment into the intercepted message. Enigma never encrypts a letter to itself which eliminates some crib positions before heavy computation. Astra wrote its own tools in Python and C++. It examined rotor orders and other parameters and checked ambiguous characters in the archival material. The search yielded a configuration with rotor order II-V-III. This order differed from known settings used that day such as V-I-II which prevented a simple reuse of a daily key. With these settings Astra produced coherent German text. The result was verified independently. The settings reproduce the decryption of all 82 characters. The key and plain text were also checked by Fred Weyerud a Enigma researcher and CryptoCellar curator. A separate analysis of MVUEH appeared in the archive. This was not a case of a model generating plausible German text and a user declaring the problem solved. There is a concrete ciphertext a concrete configuration and a reproducible result. When the decorative WWII cipher story is removed Astra did not perform any magic. It did not invent a new cryptanalysis method nor did it find a way to break Enigma instantly. The key was a classic approach find a plausible crib reduce the search space and test the remaining possibilities computationally.

Unusually a large part of this chain was assembled by a model. First archival search then matching historical documents produced a hypothesis. Under that hypothesis code was needed. The code enabled the search. The result was checked on the full message and sent for external verification. This makes the experiment more interesting than a simple model benchmark comparison. Here the LLM acted as a research agent moving between information search analysis of sources and programming depending on the next task stage. The human remained in the process Lefen posed the original task historical decrypts were created by earlier researchers and the found result was checked by an Enigma specialist. Astra linked these pieces into a workflow and took on much of the intermediate work. In my view this boundary is more interesting than how well the latest model answers in chat. Astra also helped create a project site where the search story the source data the found settings and an interactive 3D Enigma simulator are published. The simulator lets users view the result not as a dry log.