OpenAI's GPT-6 Astra Cracks an 83-Year-Old Enigma Message in Ten Hours

A Ten-Hour Crack of an 83-Year-Old Cipher

An Enigma-encrypted Wehrmacht radio message that resisted every attempt at decryption for more than eight decades has reportedly been read at last — not by a room of codebreakers, but by an artificial intelligence system working largely on its own. Carter Leffen, a product development coach at Bloomberg LP in New York, says he solved the puzzle using OpenAI's GPT-6 Astra, according to a detailed case study of the effort.

The transmission itself was tiny: 82 characters, sent in 1941, and catalogued for generations as unsolved in an archive of intercepted German army messages. According to Leffen, the configuration he calls "GPT-6 Astra Extra High" chewed on the problem for roughly ten hours, handling everything from historical research to the final verification of machine settings.

The Message That Stumped the Archive

On July 10, 1941, a German soldier used an Enigma machine to encrypt a short radio message. In plain German, he reported being in the town of Rosenow, explained that he needed instructions for his march route, and requested an immediate reply by radio.

The resulting ciphertext, tagged MVUEH, was preserved in an archive of intercepted German army messages and listed as unsolved. For 83 years nobody managed to turn those 82 characters back into readable text — a gap that stretched across the Cold War, the arrival of modern computing, and the opening years of the large language model era. The message is not historically momentous on its face. It is a soldier asking where to walk next. But it survived as a small, stubborn proof that some ciphers outlive the institutions that created them.

Why Enigma Was So Difficult to Break

The Enigma machine was one of the most important encryption devices of its era, used to protect German military traffic across theaters of the Second World War. It was first broken in 1932 by the Polish mathematician Marian Rejewski, long before the conflict began — a fact that tends to get lost in popular accounts focused on wartime codebreaking.

The scale of the problem is difficult to overstate. The machine offered roughly 159 quintillion possible daily settings, an astronomically large search space that made brute-forcing every configuration a non-starter. That was true in 1941, and it remains true for ordinary computing hardware today. Cracking an Enigma message requires exploiting structure, not exhausting possibilities.

One structural flaw proved essential to the solution. The Enigma machine never encrypts a letter as itself. If the plaintext contains an R at a given position, the ciphertext at that same position can be anything except R. That single rule dramatically narrows the space of plausible alignments, and it is exactly the kind of constraint that makes a machine-assisted search tractable.

The Crib That Unlocked the Message

Several early approaches went nowhere. Known keys from the same day of traffic did not fit the ciphertext, and automated searches returned nothing useful. Without an anchor point, even a large search space stays effectively infinite.

The breakthrough came from a second message sent on the same day that had already been decrypted. That message contained the town name "Rosenow" twice in a row. Leffen and his AI agents reasoned that the same place name might also appear somewhere in the unsolved transmission, and used it as a crib — a guessed fragment of plaintext used to test candidate settings.

Combined with Enigma's self-encipherment rule, the hypothesis eliminated a large number of positions immediately. At one position, everything lined up: the crib, the forbidden-letter constraint, and the machine logic all agreed. From there the team calculated the corresponding settings, and the remaining 68 characters fell into place as coherent German.

What the AI Agents Actually Did

The case study describes a long, multi-step workflow rather than a single prompt and a single answer. Over roughly ten hours, the system reportedly carried out a broad range of tasks:

• Searching historical archives for context, key material, and related intercepted messages
• Comparing uncertain or ambiguous letters to settle readings of the ciphertext
• Building an Enigma simulator to test candidate machine configurations
• Writing cryptanalysis code to automate parts of the search
• Running parallel experiments instead of a single linear attempt
• Testing competing keys against one another
• Cross-checking intermediate results to confirm consistency

That list matters as much as the final answer. It describes an agentic loop — plan, act, evaluate, revise — applied to a problem with a verifiable correct solution. Cryptanalysis is an unusually clean fit for such a loop, because wrong guesses fail loudly and quickly.

Reading the Recovered Text

The decrypted German contains phrases that match the soldier's apparent situation, including "Sofort Funkantwort" — immediate radio reply — and "Angabe des Marschweges," a request to specify the march route. That internal consistency, plus a message header preserved separately in the archive that confirmed the derived settings, is what the case study offers as validation.

The plaintext also carries small imperfections, likely introduced by the radio operator himself. Among them are "BTTE" standing in for "BITTE" and "WASCHBBSCH" in place of the expected spelling. Such typos are typical of hurried Morse traffic and serve as a reminder that the message was tapped out by a person under operational pressure, not generated cleanly by a machine.

What the Claim Does and Doesn't Prove

The result rests on a single developer's account and case study. No independent cryptanalytic review is described in the source material, which means the finding should be treated as a striking claim rather than a settled fact. Reproduction by outside researchers would be the natural next step.

It is also worth noting what the AI did not do. It did not discover Enigma's self-encipherment weakness, which was known to codebreakers decades ago. It did not invent the crib technique. The decisive human contribution — recognizing that a duplicated town name from the same day's traffic was worth trying as a search term — shaped the entire effort. What the system contributed was sustained execution: hours of archive digging, simulator construction, code writing, and cross-checking that would otherwise have consumed weeks of human labor.

Why It Matters Beyond One Message

The broader significance lies in what the workflow suggests about long-horizon AI tasks. Decrypting a message requires holding a goal across thousands of intermediate steps, recovering from dead ends, and resisting the temptation to declare victory on a plausible but wrong key. Those are precisely the failure modes that have limited AI agents in research settings.

If similar workflows transfer, the implications reach well beyond military history. Archives, museums, and universities hold enormous quantities of undeciphered, partially legible, or poorly catalogued material — damaged inscriptions, ambiguous manuscripts, fragmented records. Tools that can search, simulate, test, and verify at scale could make a dent in backlogs that no individual researcher could ever clear alone.

The Enigma story also carries a certain symmetry. A device built to make messages unreadable forever was first defeated by a mathematician working with paper and theory in 1932, then progressively dismantled by wartime codebreakers, and now reportedly read by a language model in ten hours. The cipher did not get weaker. The tools around it simply kept changing.

This article is based on reporting by The Decoder. Read the original article.