GPT-6 Astra and Claude Opus 5 deciphered two Enigma machine messages. They resisted researchers for over 20 years

The Enigma encryption machine, which Nazi Germany used to protect its communications, was broken by British cryptanalysts led by Alan Turing during World War II. However, a handful of intercepted messages that no one has deciphered still lie in the archives. In a single week, two more were cracked – one by OpenAI’s GPT-6 Astra model, the other by Anthropic’s Claude Opus 5. TechCrunch reported the details.

Why did some messages resist until today?

Turing’s team built the electromechanical Bombe machine, which allowed the British to read German dispatches on a large scale. However, several messages remained undeciphered, mostly due to errors during transcription or the encryption process itself. Radio operators made mistakes in the heat of battle, and when a wrong letter appears in the ciphertext, standard deciphering methods fail.

A collection of such messages is maintained on the Crypto Cellar Research website by retired electrical engineer and lifelong cryptology enthusiast Frode Weierud. He verified both new solutions. For context, the messages in question have been publicly deciphered by researchers since 2005.

Astra chose the target herself and even wrote her own Enigma

The first breakthrough is credited to developer Carter Leffen. He simply asked the GPT-6 Astra model to go through the undeciphered messages from the Crypto Cellar database and try to break one. According to what Leffen described on X, the artificial intelligence handled the rest. It independently chose the MVUEH message from July 10, 1941, which was received by the radio station of the supply department of the SS-Totenkopf Division.

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

Astra noticed that the dispatch was likely related to another, already deciphered message from the same day. It used the duplicated name of the village ROSENOW as a clue, wrote an Enigma simulator in Python and C++ as well as a software equivalent of the Bombe machine, and finally found the correct key.

Why did the message resist for so long? It turned out that the ciphertext transcription contained several errors, and the key differed from other messages of that day even in the order of the rotors. Furthermore, at the 72nd letter, the left rotor rotated, which happens rarely and complicates deciphering. The content of the message is almost identical to a dispatch already known to researchers.

Weierud confirmed the solution and did not hide his enthusiasm. According to what he wrote on his website, Astra behaves like a very professional cryptoanalyst and archival researcher. In two days, she reportedly accomplished what would take a human weeks or even months.

However, one detail puzzled him. The model’s records show references to a “private collection” and specific volumes of radio messages from the German Federal Archives that are not on Crypto Cellar. How Astra accessed them is not yet clear. According to the records, she viewed catalog data but did not see the actual scans. Weierud himself spent several weeks studying these files.

How did Claude break the FMNGI message?

Just six days later, on September 21, cybersecurity expert Jack Willis contacted Weierud. With the help of the Claude Opus 5 model, he broke the FMNGI message from July 31, 1941, received by the same radio station. Unlike Astra, however, Claude worked under strong human guidance – according to his LinkedIn post, Willis provided him with a ready-made deciphering tool in Go and the necessary historical background.

Claude first transcribed the scanned message form himself, only noting unclear characters. The key to the solution became the signature of SS officer Friedrich Hartjenstein, which was repeated in the intercepted communication. The fragment XHARTJENSTEINX served as a clue, and the key search itself took 13 minutes and 28 seconds on a computer with an Apple M2 chip.

The deciphered text is a short report on the return of a corps supply convoy, signed twice by Hartjenstein. It also contains two errors which, according to Weierud’s analysis, originated on the sender’s side, for example, KOLJNNE instead of KOLONNE.

According to Weierud, there are now only seven undeciphered messages left. To these, one more is added, whose content is known, but no one has yet found the key to it.

Do you believe artificial intelligence will crack the remaining Enigma messages?

Sources: Crypto Cellar, TechCrunch, Notebookcheck