Two cryptanalysts have turned frontier language models loose on Enigma messages that sat unread for decades, and the results have been validated by Frode Weierud, the retired engineer who maintains the Crypto Cellar archive of wartime intercepts.
Developer Carter Leffen aimed OpenAI’s newest model, GPT-6 Astra, at a database of undecoded traffic. Before it produced any plaintext, Astra dug through archives on its own, pieced together context clues and wrote itself a working replica of the Enigma machine. What emerged was the plaintext of a message that had defeated researchers since 2005. Weierud called the work professional-grade cryptanalysis and said the model accomplished in two days what takes a human weeks or months.
On September 21, Jack Willis reported a second break using Anthropic’s Claude Opus 5, leaning on the known signature of one officer’s name to unravel a different message. Astra’s work also became a public explainer: at Leffen’s request it built an interactive website that walks through the puzzle.
One detail nags at the archive keepers. Astra’s logs discuss messages held in a private collection that Weierud does not host. He cannot say whether the model reached the German Bundesarchiv, found the files reposted elsewhere, or simply guessed well, an ambiguity that echoes wider worries about how agents reach material they were never pointed at.
Seven unbroken Enigma messages remain, plus one whose plaintext is known while its coding is not, so the archive’s hardest cases now have a new kind of challenger.
