Carter Leffen from Bloomberg has reported that GPT-6 Astra successfully decoded a German military radio message dated July 10, 1941. This message had remained unsolved in the CryptoCellar archive since 2005.

Two days ago, GPT-6 Astra broke a yet unsolved German Army Enigma message from 1941. Amazingly Astra was able to autonomously:

- Search historical archives
- Compare uncertain letters
- Find contextual clues
- Build an Enigma simulator
- Write cryptanalysis code
- Run parallel… pic.twitter.com/m5usF56bsH

— Carter Leffen (@carterleffen) September 17, 2026

The radio message, identified as MVUEH, consists of 82 encrypted characters. Two known keys used on the same day proved ineffective.

The decoding process took place between September 14 and 15. Leffen set the task and guided the work while Astra, along with several parallel AI agents, examined archival materials, compared variations of poorly readable symbols, developed programs to search for the key, and tested the discovered solutions.

Leffen estimated that Astra's direct effort amounted to approximately 10 hours. However, in a detailed report, he clarified that the processes were executed in parallel, making it impossible to distinctly allocate time between the model and human contributions.

The Clue Was a Place Name

A significant hint came from another radio message dated July 10, 1941 — SIPVX, which had been decrypted in 2017. This message contained the name Rosenow mentioned twice in succession.

Researchers speculated that the sequence ROSENOWROSENOW might also appear in MVUEH. They used these 14 characters as a "crib" — a presumed fragment of the original text that helps narrow down the potential keys.

The remaining 68 characters of the message were previously unknown.

The team faced challenges due to the condition of the archival documents. Before seeking the key, they identified 12 instances where individual handwritten letters could be interpreted in multiple ways. The final reconstruction differed by three characters from one of the earlier transcriptions.

Astra developed an Enigma simulator and cryptanalysis programs in Python and C++. Under the set conditions, the system checked approximately 4.29 billion combinations of crib positions and rotor settings.

Interactive visualization of Enigma with restored settings for the MVUEH 1941 radio message. Source: MVUEH.

After identifying a candidate, around 14.8 million physical keys from the preserved header of the radio message were verified.

Ultimately, the researchers reconstructed the rotor order II-V-III, ring settings H-M-F, initial positions R-W-D, and ten connections on the plugboard.

Two independent implementations of the Enigma machine were able to decode and re-encode all 82 characters of the message and its header using the discovered parameters.

Content of the Message

The radio message requested information on the sender's further route, indicated that they were in Rosenow, and demanded an immediate radio response.

It remains unclear which specific location named Rosenow was referenced. Researchers preliminarily read the sender's signature as Waschbusch.

The result was verified by Enigma researcher Frode Weyerud, who manages the CryptoCellar archive. He confirmed the found key and noted that it significantly differed from the two other known keys from July 10, 1941.

According to him, previous attempts might have been hampered by errors in the old transcription of the radio message and an unusual rotor transition near the end of the message.

The experiment's authors emphasize that Astra did not "break the Enigma" as a cipher. The methods of its cryptanalysis existed before and during World War II. In this instance, the researchers restored the key and text of a specific radio message that had remained unsolved in a modern archive.

They also do not claim to have achieved the first-ever decryption of such a message: the search was comprehensive only within the given assumptions about the type of machine, rotors, plugboard settings, variations in reading the original document, and the chosen crib.

It is worth noting that earlier in September, OpenAI announced a claimed solution to the problem of existence and smoothness of the Navier-Stokes equations — one of the seven Millennium Prize Problems for which the Clay Institute has set a $1 million prize.