Anthropic’s Claude Helps Solve a 1941 Enigma Message

A message sent to a German military radio station in 1941 has yielded its Enigma key and plaintext. Anthropic’s Claude Opus 5 helped Jack Willis solve the message, known as FMNGI, and cryptography archivist Frode Weierud checked the result against an outgoing copy preserved in the archives.

How they reached that result matters. Willis supplied Claude with a cryptanalysis workbench written in Go and historical material that led to a promising name-based crib, or suspected fragment of plaintext. Weierud describes the attack as strongly guided by a human. It is a documented AI-assisted solution, not evidence that Claude can routinely break Enigma messages on its own.

A second recent break offers a useful contrast. Weierud says OpenAI’s GPT-6 Astra took more initiative in selecting and developing an attack on a different message, MVUEH. Together, the cases show what these models can contribute to historical research and why the conditions behind a successful result matter.

An Archival Message Provides a Check on the Solution

FMNGI was an incoming Enigma message dated July 31, 1941, and recorded by the quartermaster’s radio station of the SS-Totenkopf Division. Its received message form identifies the sender only by a tactical radio callsign. Crypto Cellar Research had published a transcription in 2005, but the message remained unsolved.

In Weierud’s account of the FMNGI break, Willis reported that Claude’s search recovered an Enigma key and this plaintext:

KOLJNNEVONXKORPSNACHSCHUBZURUEKXHARTJENSTEINXHARTJENSTEINX

Enigma messages did not preserve spaces, so interpreting that sequence requires more than inserting word breaks. Weierud reads its opening as an imperfect rendering of Kolonne von Korpsnachschub zurück, followed by the repeated name Hartjenstein. The odd-looking letters are part of the evidence: KOLJNNE contains an error, and Weierud says ZURUEK should have been encoded as ZURUEQ under the operators’ convention of replacing CK with Q.

Weierud could compare the recovered text with an outgoing archival message, No. 205 NF. That provides a check beyond a phrase that merely looks plausible in German. The comparison supports the reported recovery while revealing mistakes made around the message’s original preparation or transmission.

The archival record also calls for a qualification. The outgoing collection, which included material related to FMNGI, had been found and added to Crypto Cellar Research’s message resources in July 2026. This was not the recovery of text for which no corresponding archival record existed anywhere. Weierud says he thought to consult the full plaintext after Willis reported the break; his account of Claude’s attack instead centers on ciphertext transcription, related traffic and a shorter crib.

Claude Worked With a Specialist’s Workbench

Willis did not give Claude an unexplained string of letters and ask it to invent an attack from scratch. According to Weierud, he provided a Go cryptanalysis workbench and historical material, then directed the model to investigate FMNGI. Software for testing Enigma settings was part of the cryptanalytic capability in this case, and Willis supplied it.

The historical preparation mattered too. Willis initially lacked the copy of FMNGI’s received form that Weierud later published. The received text had long been available as a transcription, but handwritten radio records can introduce uncertainty of their own. Radio reception, Morse transmission and transcription can all affect the letters a researcher tries to decrypt.

Claude also had access to the outgoing form associated with the message. Weierud says it transcribed that scanned form and recorded ambiguous characters rather than silently choosing one reading. Another already-solved message, ALQFI, had traveled between the same two radio stations and provided a way to check the transcription work.

Weierud dates the successful final key search to September 20, 2026. Willis reported that it took 13 minutes and 28 seconds on an Apple M2 host; he emailed Weierud about the break the next day. That figure describes the final search. It excludes the time required to assemble the archive, prepare the tools, identify a target and investigate possible clues, so it should not be taken as the time Claude needed to solve the entire research problem.

A Repeated Name Made the Search More Promising

The decisive clue was a likely signature. ALQFI ended with the name of Friedrich Hartjenstein, and other traffic gave researchers reason to expect the same name in FMNGI. The resulting 14-character crib was XHARTJENSTEINX, with X used as a separator.

A crib is a suspected piece of plaintext that can be tested against ciphertext. It does not supply the Enigma settings by itself, but it narrows a search that would otherwise have to consider far more possibilities. In FMNGI, the recovered text contains Hartjenstein’s name twice, showing why that lead was productive.

Weierud notes that Hartjenstein’s signature had already proved useful in earlier cryptanalysis, with several documented variations. The work here lay in bringing a plausible name, the relevant messages and a working search environment together, then finding settings that produced the reported plaintext. It was informed historical inference, not a model conjuring a key from general knowledge.

Even a good crib must survive messy source material. The first word’s error shows why a rigid demand for perfectly copied German could have led an investigation astray. The solution had to accommodate the message as it was enciphered and recorded, rather than an idealized version of what its sender meant to write.

GPT-6 Astra’s MVUEH Break Involved More Initiative

Six days before Willis contacted him, Weierud received another request to validate an AI-assisted Enigma solution. This one concerned MVUEH, a message from July 10, 1941, that had resisted solution since 2005. Carter Leffen had asked GPT-6 Astra to see whether it could break any of the unsolved messages on Crypto Cellar Research’s site.

Weierud’s MVUEH account describes a different division of labor. He says Astra chose MVUEH as a promising target, noticed its possible connection to another message from the same day, developed Enigma software in Python and C++, and pursued a crib based on the repeated place name ROSENOW. Willis, by contrast, gave Claude an existing Go workbench and a more directed FMNGI investigation.

MVUEH’s validation had a strong cross-check. Its recovered plaintext was almost identical to that of the previously solved message SIPVX, though the messages had different lengths and used different Enigma settings. Weierud says the reported MVUEH key and plaintext were immediately convincing when Leffen sent them. Their close relationship also helps explain how Astra found a useful lead; MVUEH was not a wholly isolated ciphertext without context.

How independently Astra conducted every step remains a narrower question than whether MVUEH was solved. Weierud said he was continuing to examine the model’s logs. One passage named Bundesarchiv file references not available on his website, but he could not establish from that passage whether Astra had accessed digitized archival scans or found the references elsewhere. Its apparent research path should not be presented as fully audited.

In both cases, a human defined the broad problem and an expert assessed the result. Within those boundaries, Weierud credits Astra with more of the target selection and tool-building, while Claude’s FMNGI solution depended on more substantial specialist preparation. Calling both simply “AI breaks Enigma” erases that distinction.

A Verified Plaintext Is Not a General Encryption Break

FMNGI and MVUEH are specific solutions to specific historical messages. For FMNGI, the reported key yields a plaintext that Weierud checked against an outgoing archival copy. For MVUEH, the reported key and plaintext fit a closely related message that was already solved. These are meaningful results, not just model-generated guesses that sound like wartime German.

They do not establish a success rate on other unsolved traffic. Weierud notes that his collection still had seven unbroken Enigma messages, along with a separate puzzle involving a message believed to share another message’s plaintext. The two successes do not tell us whether the remaining cases have similarly useful cribs, records or relationships to solved traffic.

Nor do they demonstrate an ability to defeat modern encryption. These investigations concern historical Enigma traffic, where archival context, suspected plaintext, machine-specific software and surviving message copies can make a targeted search possible. Nothing in either documented result tests a modern cryptographic system.

FMNGI shows a model helping turn specialist tools and imperfect historical evidence into a checked solution to a message that had resisted earlier work. To judge the next AI-assisted discovery, the question is not only whether the answer is right, but also what the model was given and which parts of the research it actually performed.