AI Wins › Science & health
Amateurs used AI to crack Enigma messages unsolved since 2005
High verification: Named in Nature reporting by Kaia Glickman, with independent confirmation from cryptanalyst Frode Weierud of Crypto Cellar Research, who had kept the messages on his unsolved list. How we check stories
What happened
In mid-September 2026, Carter Leffen, a product development coach at Bloomberg who is not a cryptographer, pointed OpenAI's GPT-6 Astra at a public list of unbroken World War II Enigma messages kept by Frode Weierud at Crypto Cellar Research1. The message it picked, catalogued as MVUEH, was an 82-letter German army radio message received on July 10, 1941, and it had resisted every attempt to break it since 2005.
According to Weierud, Leffen only directed the model to see whether it could break any of the unsolved messages on the site. Astra chose MVUEH, guessed that its text might be related to an already-broken message from the same day, and used the repeated place name ROSENOW as a known-plaintext "crib." It then wrote Python and C++ software for an Enigma simulator and a search program modelled on the wartime Bombe machines used to find Enigma keys, and recovered a key and plaintext. Weierud, who has worked this collection for years, wrote that "it was immediately clear he had found the correct key and plaintext."
The recovered text is a short operational request: a march route, a location at Rosenow, and an immediate radio reply. Weierud noted that the key was different from other keys used that day, that the original ciphertext transcript had several errors, and that a rare left-hand wheel turnover late in the message had likely blocked earlier attacks. He also wrote that the work, done in two days, would have taken a human researcher weeks or months.
Kaia Glickman reported the story for Nature2 on October 16, 2026, describing Leffen as having no cryptography experience (Nature first misstated his employer and later corrected it to Bloomberg). Bruce Schneier also carried Weierud's account3 on September 22.
A second break, with Claude
Six days later, Weierud posted that Jack Willis4 had broken another unsolved message from the same collection, FMNGI, dated July 31, 1941. This time the tool was Anthropic's Claude Opus 5. Weierud said the attack was more directed: Willis gave Claude a cryptanalysis workbench and historical material. Claude used a likely signature as a crib. Willis reported that the final key search ran for about 13 and a half minutes on an Apple M2 computer, on September 20, 2026. Weierud confirmed the plaintext.
Other remaining messages in the collection were also broken later that month. Weierud published John Woodhouse's break of two short, heavily garbled messages, AWTZK and ZNLZT, after checking the decipherment; that write-up does not name an AI model, so we don't attribute it to one.
Why this matters if you're choosing an AI plan
- Hard research plus writing search software is where top reasoning models show their value, not everyday chat.
- An independent expert still had to confirm the key and plaintext. Without that check, it would only be a claim.
- This isn't a one-model story: Claude Opus 5 also broke a later message in the same collection, with more human guidance.
- A mid-priced plan may not be enough if you need the highest reasoning settings or long unattended runs.
Credits. Reporting by Kaia Glickman for Nature. Primary validation and technical accounts by Frode Weierud of Crypto Cellar Research. Further notice by Bruce Schneier. The first break is by Carter Leffen using GPT-6 Astra; the FMNGI break is by Jack Willis using Claude Opus 5. Up to Speed with AI has no relationship with the people or companies named, and nothing here suggests they endorse us or any AI plan.
Spotted an error, or are you in this story and want something changed? Tell us and we'll fix or remove it.
More AI wins
A couple used AI to help get a $22,604 surprise ER bill waived
Jackie Davalos got a collections notice for an ER visit she never billed for. Her partner used Claude to help dispute it; the hospital later waived the whole amount.
One developer, one AI model, and free look-alikes of Photoshop, Premiere and Excel
A veteran engineer used Claude to build free, open-source stand-ins for Adobe and Microsoft Office apps. Real and downloadable, but early alpha.
A 23-year-old amateur used AI to help crack a 60-year-old math problem
An amateur gave an unsolved Erdős problem to GPT-5.4 Pro. Terence Tao and Jared Duker Lichtman checked the proof and say the key idea is new.
A family says a $20 AI plan helped cut a $195,000 hospital bill to about $33,000
A grieving family used Claude to decode billing codes, flag duplicate charges and write dispute letters. Self-reported, so read it that way.
A tech writer built his own stand-ins for $400 a year of app subscriptions, using a $20 AI plan
How-To Geek's Dibakar Ghosh used Claude to build home-made swaps for a grammar checker, a dictation app and a time tracker. Real savings, with a few catches.
