GPT‑6 Astra cracks Enigma MVUEH message from 1941

GPT‑6 Astra solves a 75‑year‑old Enigma cipher

Takeaway: On 15 September 2026 OpenAI’s GPT‑6 model, codenamed Astra, independently broke the German Army Enigma message MVUEH (10 July 1941), a ciphertext that had remained unsolved since 2005, by discovering a novel key, wheel order, and a rare left‑rotor turnover.

Why the MVUEH message mattered

The MVUEH ciphertext was logged as message Nr. 172 by the SS‑Totenkopf Quartiermeister, Ib, on 10 July 1941 and has been part of the Crypto Cellar Research “1941 Message List” since its publication.
All other traffic from that day used the daily Enigma key with wheel order 512. MVUEH required a completely different wheel order (253) and distinct plugboard and ring settings, a fact that misled earlier analysts.
The message length (82 characters) and a transcription error (“Btte” instead of “Bitte”) further complicated traditional crib attacks.
A rare left‑hand rotor turnover at the 72nd character was present; such turnovers are infrequent and break many automated cryptanalytic heuristics.

These anomalies explain why the message resisted solution for over two decades.

How GPT‑6 Astra approached the problem

Prompt and goal selection: Carter Leffer instructed Astra to scan the Crypto Cellar website for any unsolved Enigma messages. The model identified MVUEH as the most promising target and noted a possible relationship to the already‑solved message Nr. 173 (SIPVX).
Crib generation: Astra hypothesized that the repeated place name ROSENOW ROSENOW could serve as a crib, a common technique in Enigma cryptanalysis.
Tool creation: The model wrote Python and C++ code to implement an Enigma simulator and a Bombe‑style search engine, integrating the crib into the search space.
Search execution: Using the generated software, Astra performed a systematic search over wheel orders, plugboard configurations, and ring settings, eventually discovering the unique key (wheel order 253) and confirming the left‑rotor turnover.
Verification: Re‑encryption of the recovered plaintext with the discovered settings reproduced the original ciphertext byte‑for‑byte, confirming the correctness of the break.

The recovered plaintext and its significance

Ciphertext (as transcribed):

BTTE UM ANGABE DES MARSQWEGES X BEFINDE MIQ IN X ROSENOW ROSENOW X SOFORT FUNKANTWORT X WASCHBBSCH

Plaintext after correction:

BITTE UM ANGABE DES MARSCHWEGES. BEFINDE MICH IN ROSENOW, ROSENOW. SOFORT FUNKANTWORT. WASCHBUSCH.

English translation:

“Please specify the marching route. I am in Rosenow, Rosenow. Immediate radio reply required. Waschbusch.”

The plaintext is almost identical to that of message Nr. 173 (SIPVX), differing only by twelve characters caused by the typo and a duplicated signature.

Insights from the AI’s research process

Astra located references to Bundesarchiv radio‑message collections (volumes RS 3‑3/20a and RS 3‑3/63b) that are not publicly listed on the Crypto Cellar site. The model’s internal “evidence pass” traced the corpus to a private collection, suggesting it can autonomously discover archival leads.
The AI’s logs show a two‑day workflow: (1) data gathering and hypothesis formation, (2) software generation and exhaustive key search. Human oversight was limited to the initial prompt and later verification of the output.
Commenters noted that other LLMs (e.g., Qwen 3.7, Gemini 3.8 Flash) can also decode the ciphertext when given the same crib, but Astra uniquely performed the end‑to‑end pipeline without external tooling.

Community reactions and skepticism

Human‑in‑the‑loop importance: Some commenters emphasized that Leffer’s direction was crucial, arguing that the breakthrough still relied on human insight to select the target and interpret the results.
Software originality: Questions were raised about how much of the generated simulator code was novel versus repurposed from existing open‑source Enigma tools.
Verification concerns: A few users asked how certainty about the plaintext is established, noting that Enigma’s many‑to‑one mappings can produce plausible but incorrect decodings.
Broader implications: Several participants linked this success to the growing ability of LLMs to assist in cryptanalysis, speculating about future applications to unsolved historical ciphers (e.g., Zodiac, Kryptos) and even the creation of new encryption schemes.

What this means for cryptanalysis

Automation of classic attacks: GPT‑6 demonstrates that large language models can not only suggest cribs but also generate and execute the full cryptanalytic pipeline, reducing weeks of manual work to days.
Discovery of hidden archival material: The model’s ability to locate obscure references suggests future AI‑assisted archival research could accelerate historical cryptanalysis.
Human‑AI collaboration model: While the AI performed the heavy lifting, human expertise remained essential for prompt design, result validation, and contextual interpretation.

Future directions

Benchmarking against untouched ciphers: Researchers should test Astra on Enigma messages with no prior solutions to assess genuine novelty.
Open‑source reproducibility: Publishing the generated simulator code and search logs would enable independent verification and foster community improvements.
Ethical considerations: The ability of LLMs to locate and exploit private archival collections raises questions about data provenance and responsible use.

The break of MVUEH showcases the emerging power of large language models to augment and, in some cases, automate complex cryptanalytic tasks that have stumped experts for decades.