Astra and Claude Opus 5 models help decipher two long-unsolved Enigma messages
Cryptanalysts used the Astra (OpenAI) and Claude Opus 5 (Anthropic) models to decipher two long-unsolved Enigma messages, one undeciphered since 2005. Seven messages remain undeciphered.
Two cryptanalysts announced that, using language models, they managed to decrypt two long-unsolved messages encrypted by the German Enigma machine during World War II. Developer Carter Leffen tasked OpenAI's Astra model with finding an undeciphered message in a database and decoding it. According to the article, the model carried out archival research on its own, searched for contextual clues, built its own Enigma machine simulator, and obtained the plaintext of a message that had remained undeciphered since 2005.
The solution was validated by Frode Weierud, a retired electrical engineer who runs the Crypto Cellar archive containing a database of still-undeciphered Enigma messages. According to him, the work that the Astra model completed in two days would have taken a human weeks to months; he himself spent several weeks researching the same archival materials from the Bundesarchiv. While working, the model also mentioned messages from a "private collection" that is not part of the public database – Weierud is not sure whether the model actually had access to them or whether it was a different source.
A few days later, on 21 September 2026, another cryptanalyst, Jack Willis, contacted Weierud saying that using Anthropic's Claude Opus 5 model, he had deciphered another unsolved message. Unlike the case with the Astra model, Willis provided the model with substantially more clues, including a known signature pattern of a specific German officer, which helped the model crack the message.
According to Weierud, seven Enigma messages remain undeciphered, plus one more for which the plaintext is known but the encryption key itself is not yet known.
Why it matters
The case demonstrates the ability of current models to independently manage a multi-day research process – from archival research through building a simulation to the actual solution – without detailed human instructions, a type of task previously considered the domain of narrowly specialized experts. It also shows a difference between models: one worked practically autonomously, the other needed significantly more human guidance, which is a relevant clue when deciding how much oversight similar tasks require.
Two audiences, two different impacts
What this means
For individuals
The difference in approach is notable: the Astra model worked with minimal input and sourced the context itself, while the Claude Opus 5 model needed more concrete guidance from a human – so when assigning similar AI research tasks, it's worth considering how much initiative to expect from the model.
For a business
The case shows that current state-of-the-art models can autonomously carry out multi-stage archival research, build simulations, and solve tasks without a clearly defined procedure – a relevant signal for companies considering deploying AI agents on complex analytical and research tasks where no ready-made methodology exists.
Development More business impacts →Check the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.