# An AI-assisted Enigma break began with a clue in another message

> An AI-assisted Enigma recovery connects archival clues with a search others can inspect. The case offers practical lessons for research with agents.

By BIG CHANGE Editorial

Published: 2026-09-22T15:53:42.087Z
Updated: 2026-09-22T16:08:50.868Z
Canonical: https://bigchange.ai/blog/ai-enigma-mvueh-archive-break

![Three exposed cipher-like rotors beside a matching notched fragment, with an orange line passing through the assembly.](https://bigchange.ai/api/media/file/archival-clue-rotor-search-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE. A related clue can narrow a cipher search; the resulting answer still needs independent checks. This conceptual rotor assembly illustrates that relationship, not the historical machine or its recovered settings.

The newly recovered message asks for a marching route and an immediate radio reply. [The published investigation](https://mvueh-enigma-solved.carterl.chatgpt.site/)

Known as MVUEH, the Enigma message dates from July 10, 1941. Archive researcher Frode Weierud says it resisted attempts since 2005. He received a solution on September 15, 2026, and validated its key and plaintext. [Weierud’s September 19 account](https://www.cryptocellar.org/bgac/the-mvueh-break.html)

The satisfying part of this story is that an unreadable record now contains something a historian can use. The instructive part is how the search got there. A clue in previously solved traffic suggested what to look for; explicit rules constrained the search; a recovered machine configuration made the answer open to inspection.

## The big change

- **What changed:** AI agents helped turn a historical clue into software that tested possible answers. The broader shift is that researchers can delegate parts of an investigation, including building its tools, while retaining responsibility for the question and the evidence.
- **Why it matters:** Museums, archives and research teams could pursue leads that would otherwise wait for scarce time and specialist programming. This case gives no evidence that modern encryption or today’s passwords have been broken.
- **What to watch:** A plausible next step is research software that links collections, proposes leads and tests them. If those tools prove dependable, specialists can spend less time constructing each search and more time interpreting and challenging its results. Institutions with searchable records and clear source histories may benefit first.

## A neighboring message supplied a foothold

Enigma was an electromechanical cipher machine. Its wired rotors and plug connections transformed letters, with the mapping changing as the machine operated. Recovering an individual message meant finding settings that could undo those transformations. The Allies were reading Enigma traffic during the war; this is a newly solved archival message. [The NSA museum’s explanation](https://www.nsa.gov/History/National-Cryptologic-Museum/Exhibits-Artifacts/Exhibit-View/Article/2719176/world-war-2-enigma/)

The archive identifies an 82-letter encrypted body for MVUEH and lists both incoming and outgoing records. A related message, SIPVX, had already been solved by Alex Shovkoplyas in June 2017. The collection also grew after research in Germany’s Bundesarchiv located additional radio-message material in July 2026. The ingredients available to a new investigator therefore included earlier human successes and additional archival work. [The message inventory](https://cryptocellar.org/bgac/1941-msg-list.html)

According to the solver’s account, the repeated place name ROSENOWROSENOW in SIPVX became a 14-letter candidate for MVUEH. The investigation used GPT-6 Astra and specialist agents to develop and run searches. A proposed fragment of the original message gave that computation a target. [The investigation account](https://mvueh-enigma-solved.carterl.chatgpt.site/)

Cryptanalysts call such a fragment a *crib*. It lets them ask whether a particular configuration could turn those proposed letters into the observed ciphertext. Enigma’s property that a letter cannot encrypt to itself supplies one way to reject an impossible alignment. The National Museum of Computing explains how the wartime Bombe used cribs and machine constraints to eliminate possibilities while searching for settings. These principles long predate language models. [The museum’s Bombe explanation](https://www.tnmoc.org/bombe)

A crib carries a risk as well as an advantage. Guessing a place name directs effort toward a tractable question, but success has to produce evidence beyond the guessed name. Otherwise the researcher has simply found a way to recover the words supplied at the start.

Software can examine far more configurations than someone can inspect by hand. An agent can also help prepare that software from a research hypothesis. Connecting a historical clue to an executable test becomes useful when the test retains enough independence to reject a bad idea.

## Uncertain letters need to remain uncertain

Historical records complicate an otherwise exact calculation. A cipher solver receives a transcription, and a transcription is already an interpretation of a document.

The published transcription report says the investigation preserved a finite set of alternatives for uncertain letters before recovering the key. The preferred reading stayed within those alternatives, and the original source strings remained available. It also distinguishes the literal decipherment from later spelling repairs and a tentative interpretation of the signature. Those editorial changes were excluded from the exact cryptographic check. [The transcription report](https://mvueh-enigma-solved.carterl.chatgpt.site/downloads/transcription.md)

That discipline has a practical use beyond ciphers. Imagine an archivist working from a faint shipping ledger. One character might be an N or an M. Keeping both readings allows a later comparison with another record to help resolve it. Quietly choosing the convenient letter after finding a matching destination conceals how much freedom the researcher used.

An AI system asked to produce clean, readable text can make this problem harder to notice. A polished transcription may hide the exact place where evidence was weak. For research, the useful output could be an awkward record of alternatives, with links back to the source, followed by a separate readable interpretation.

That separation also makes disagreement productive. Another researcher can challenge the letter without having to discard the entire account. An archive can retain the original image, the competing readings and the proposed interpretation as distinct records. The result remains open to correction when a better scan or a related document appears.

## A plausible sentence needs more than a round trip

The completion report makes an easily missed point: decrypting text and then encrypting it back is insufficient evidence by itself. Enigma is reversible, so even a wrong key can round-trip its own output. The report instead assembles several checks: readable language outside the supplied crib, consistency with the message header, source-constrained letter alternatives and agreement between implementations. These are the investigators’ reported checks; we have not rerun their programs. [The completion report](https://mvueh-enigma-solved.carterl.chatgpt.site/downloads/completion-report.md)

The archive’s recovered-key table supplies concrete settings, including rotor order II, V, III and starting position RWD. A claim presented this way offers something much more specific to dispute than a model’s confident translation. [The published July key table](https://cryptocellar.org/bgac/e-keys-july-1941.html)

Consider the difference from asking an assistant to infer what a damaged letter probably said. A convincing reconstruction might fit the surrounding history while leaving many alternative sentences possible. Here, a proposed configuration also has to obey the cipher’s mechanical rules. Those constraints give a reviewer additional ways to catch mistakes.

Even then, checks have scopes. The completion report explains that the header alone leaves multiple possible keys and that the search’s completeness depends on its declared machine model and permitted transcription alternatives. A later experiment used a phrase learned from the recovered answer; it was a follow-up check, not another blind discovery. [The search’s stated limits](https://mvueh-enigma-solved.carterl.chatgpt.site/downloads/completion-report.md)

For anyone commissioning AI research, this suggests a useful distinction between generating a candidate and deciding to accept it. The acceptance test should expose what the candidate had to satisfy, which assumptions were supplied and which observations remained available to test the answer. A long activity log can help reconstruct the work, but its length is no substitute for a check that could have failed.

## The answer and the autonomy claim have different evidence

Weierud describes the AI as solving the problem entirely on its own, with log analysis still underway. [The archive account](https://www.cryptocellar.org/bgac/the-mvueh-break.html)

The solver describes researcher-led work with parallel AI agents over September 14 and 15, without a complete tally of human or model reasoning time. [The participant’s process description](https://mvueh-enigma-solved.carterl.chatgpt.site/)

These accounts support an AI-assisted recovery. Establishing autonomy requires fuller evidence of human interventions and decisions.

Nor does the long period when the message remained unsolved measure the time a comparable human investigation would need with the same records and tools. Calendar years in an archive are not continuous researcher-hours. A fair speed comparison would have to account for the starting materials, effort and failed attempts on both sides.

This leaves plenty to be interested in. A researcher who can enlist agents to investigate clues and build purpose-specific search programs may be able to attempt work that would otherwise remain on a list. The relevant benefit could be more worthwhile investigations reaching a testable result, even where human direction remains substantial. Measuring that benefit requires multiple cases, including the ones that fail.

## The opportunity is in the next archive question

Crypto Cellar’s project history describes work stretching back to the early 2000s, with researchers publishing recovered keys and grappling with difficult source material. The current result depends on that accumulated record being available to investigate. [The archive project’s history](https://cryptocellar.org/bgac/index.html)

For an archive with limited staff, a promising experiment would start with a bounded collection and a question whose answer can be checked against something else. An agent might propose links between records or prepare a search over possible readings. The archivist would need the rejected possibilities and unresolved ambiguities as well as the attractive matches. Publication would distinguish a recovered fact from a suggested connection.

The optimistic prospect is that more collections receive that attention. The less comfortable possibility is that inexpensive searching produces a large queue of convincing proposals that specialists must laboriously disprove. The balance depends on whether each proposal arrives with useful evidence and whether the institution has capacity to examine it. Counting generated interpretations would tell an archive little about how much reliable knowledge it had gained.

There is also a clear technical boundary. Modern AES, for example, is a block cipher specified by NIST, operating on 128-bit blocks with defined key lengths. Its structure is different from Enigma’s rotor machine. Nothing in this investigation demonstrates an attack on AES or establishes that contemporary encryption has generally become breakable. [The AES standard](https://csrc.nist.gov/pubs/fips/197/final)

The precise Rosenow location and a tentative signature remain open in the case study. Recovering letters does not automatically identify the place or person behind them. [The remaining historical questions](https://mvueh-enigma-solved.carterl.chatgpt.site/)

Those questions give the next investigator somewhere concrete to begin: a readable message, a proposed interpretation and an explicit account of what still needs evidence.

## Sources

- [Crypto Cellar: The MVUEH break](https://www.cryptocellar.org/bgac/the-mvueh-break.html) — Weierud reports validating the solution. His autonomy characterization differs from the participant’s detailed account; log analysis was ongoing.
- [MVUEH cryptanalysis case study](https://mvueh-enigma-solved.carterl.chatgpt.site/) — Participant account describes researcher-led work, agents and surviving historical questions. Its two-day timeline is not a controlled speed comparison.
- [NSA museum: World War II Enigma](https://www.nsa.gov/History/National-Cryptologic-Museum/Exhibits-Artifacts/Exhibit-View/Article/2719176/world-war-2-enigma/) — Explains the machine and wartime codebreaking context. It establishes historical mechanics, not the present investigation’s performance.
- [Crypto Cellar: 1941 message inventory](https://cryptocellar.org/bgac/1941-msg-list.html) — Lists MVUEH’s incoming and outgoing records, SIPVX’s earlier solution and additional archival material. Links lead to the message forms.
- [National Museum of Computing: The Bombe](https://www.tnmoc.org/bombe) — The museum operates a reconstructed Bombe and explains cribs and machine constraints. These methods predate AI agents.
- [MVUEH transcription report](https://mvueh-enigma-solved.carterl.chatgpt.site/downloads/transcription.md) — Detailed record of permitted letter alternatives, preserved source strings and editorial normalization. This is the investigators’ own account.
- [MVUEH completion report](https://mvueh-enigma-solved.carterl.chatgpt.site/downloads/completion-report.md) — Documents verification arguments, bounded search assumptions and later checks. We read this report without reproducing its computation.
- [Crypto Cellar: July 1941 recovered keys](https://cryptocellar.org/bgac/e-keys-july-1941.html) — Publishes the recovered MVUEH machine settings. A specific key is available for inspection; this article does not certify a fresh rerun.
- [Crypto Cellar: Breaking German Army Ciphers](https://cryptocellar.org/bgac/index.html) — Project history documents earlier research and the publication of archival material. Some historical status language predates the latest individual solutions.
- [NIST: Advanced Encryption Standard](https://csrc.nist.gov/pubs/fips/197/final) — Defines AES’s block structure and key lengths. The Enigma investigation does not test or break this modern cipher.