AI just broke a historic Enigma message

[Jürgen]

If you are into this kind of thing -- an Enigma message from 1941 that was still unbroken has been cracked last week by OpenAI's GTP6-Astra. There's a nicely explained (although rather long and somewhat overstyled) report here:  https://mvueh-enigma-solved.carterl.chatgpt.site/ 

The message was rather short (82 letters), making it hard to decrypt. The eventual key to success was guessing a strong crib, the repeated place name "Rosenow Rosenow". From the account above, it seems that this lucky guess was still made by a human operator, not by the AI model.

10 hours or so of model time; they don't spell out the cost... (Although it can apparently be derived from the audit file available for download on that page.) 

[Jürgen]

Frode Weierud just reached out to me with additional detail. The OpenAI team had contacted him right after achieving the break, asking him to validate their results. Frode confirms that the entire break was performed autonomously by GPT6-Astra, including the identification of the crib. That makes the effort even more impressive in my view -- the AI model had found and pulled in a lot of context, some of it from yet-to-be clarified sources.

Frode has published a short summary here: https://cryptocellar.org/bgac/the-mvueh-break.html.