OpenAI's newest flagship model, GPT-6 Astra, has accomplished what human cryptographers could not: breaking a German Army Enigma message from July 10, 1941, that had remained unsolved for 21 years. The 82-letter message, known by its indicator MVUEH, was confirmed decrypted by Frode Weierud, a leading Enigma researcher, who verified the breakthrough on September 15.

What Makes This Enigma Break Historically Significant?

The message had been listed among unbroken Enigma intercepts on Weierud's CryptoCellar research site since 2005, resisting all attempts to crack it. The plaintext reads: "Please specify the route of march. I am in Rosenow, Rosenow. Immediate reply by radio," signed by a sender tentatively identified as Waschbusch. The key used Enigma I with reflector B, rotor order II-V-III, ring settings H-M-F, and ten plugboard pairs, which was completely different from other keys used on July 10, 1941.

Two major obstacles had prevented earlier solutions. The published transcription of the ciphertext contained several errors, and the machine's left-hand rotor turned over at the 72nd letter, a rare event that significantly complicates cryptanalytic attacks.

How Did GPT-6 Astra Solve the Unsolvable?

The breakthrough involved a combination of AI autonomy and human guidance. Carter Leffen, who ran the model, used a crib (a suspected piece of plaintext) from a related message, SIPVX, which had been broken in 2017 and contained the repeated town name ROSENOWROSENOW. However, accounts differ on how much the model accomplished independently.

Leffen's case study describes the work as "a researcher-led investigation with GPT-6 Astra and parallel specialist agents," in which the researcher set goals and pushed the investigation forward. However, Weierud's account credits the model far more substantially. According to Weierud, Leffen only asked the model to try the unbroken messages on his site, and GPT-6 Astra independently picked MVUEH, spotted the link to SIPVX, chose the crib, and wrote its own Enigma simulator and Bombe software in Python and C++.

"As an old cryptanalyst, I am still in awe," stated Frode Weierud, Enigma researcher at CryptoCellar.

Frode Weierud, Enigma Researcher at CryptoCellar

Leffen's own post on X (formerly Twitter) aligns more closely with Weierud's assessment, listing searching historical archives, comparing uncertain letters, building an Enigma simulator, and recovering the plaintext among the things the model did "autonomously".

What Technical Capabilities Did GPT-6 Astra Demonstrate?

• Cryptanalytic Reasoning: The model identified patterns across historical messages and recognized connections between the unsolved MVUEH message and the previously broken SIPVX intercept from the same day.
• Software Development: GPT-6 Astra wrote functional implementations of both an Enigma simulator and Bombe software (the machine used to break Enigma codes) in multiple programming languages without human intervention.
• Verification and Validation: The model conducted exhaustive searches, accounting for all 43,016 batches of the Rosenow search to confirm the solution's correctness.

Verification remains the strongest aspect of this breakthrough. Weierud, who received the details on September 15, confirmed that "it was immediately clear he had found the correct key and plaintext." Separate implementations reproduced both the full message body and its header, providing independent confirmation of the solution.

The case study is careful about its claims, stating: "This is not the original breaking of Enigma, and no first-ever or corpus-wide breakthrough is claimed." The work was conducted by a user rather than announced as an official OpenAI achievement. OpenAI's GPT-6 Astra launch materials pitch the model on computer use, science, and cybersecurity capabilities, and the model is available to ChatGPT Plus, Pro, Business, and Enterprise users as well as through the API.

The breakthrough comes as OpenAI continues to advance its flagship models. In July, CEO Sam Altman defended language models against critics who called them a dead end. This Enigma decryption demonstrates that modern large language models (LLMs), which are AI systems trained on vast amounts of text data, can apply reasoning and coding skills to historically difficult problems in ways that extend far beyond their original training.