Astra’s route through the archive
Developer Carter Leffen asked Astra to find an unsolved Enigma message in a database and decipher it. The model searched archival material for contextual clues, created a simulator of the Enigma machine and reconstructed the message’s plaintext. Leffen then used Astra to build an interactive site explaining the problem.
Astra’s work also left an unresolved question about its sources. Its logs refer to messages from a “private collection” that is not on Frode Weierud’s Crypto Cellar site. Weierud, a retired electrical engineer who has spent his life studying cryptology, does not know whether Astra had access to those materials. He suggested another researcher may have put them online, or that the model may have used public German government archives.
Weierud checked Leffen’s solution last week and said he was astonished. In his assessment, GPT–6 Astra behaved like a professional cryptanalyst and archival researcher, doing in two days work that might take a person weeks or months. Weierud himself had spent several weeks examining files from Germany’s Federal Archive cited by Astra.
A second model, a different clue
On September 21, cryptanalyst and cybersecurity executive Jack Willis told Weierud that he had deciphered another unsolved message using Anthropic’s Claude Opus 5. Unlike Leffen, Willis gave the model more guidance: Claude used a known signature bearing a specific officer’s name to reveal the message.
The two results show different kinds of assistance. Astra’s account describes a model finding material and assembling a method; Claude’s relied on a supplied clue. That makes the remaining uncertainty about Astra’s sources more than a detail: it is difficult to judge how much independent archival work the model did without knowing what material it could reach.
Seven Enigma messages remain undeciphered, Weierud said. There is also one more message whose plaintext is known, but whose code has not been solved. The models have narrowed the archive’s mysteries, not closed the case.
Daily AI news
Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.
Only what matters — every day
Follow on X