GPT-6 Astra Cracks an Old Enigma Message on Its Own

Cryptographer Bruce Schneier flags a curious episode: the GPT-6 Astra system managed to break one of the old, still-unread Enigma messages. The most striking part is that it did so with almost no human involvement.

Direction, not instructions

Carter Leffer only handed GPT-6 Astra a broad task — see whether it could crack any of the unbroken Enigma messages published on the Crypto Cellar Research site. Everything else the system worked out for itself.

After surveying the available intercepts, the model judged message No. 172, designated MVUEH, to be the most promising. It also quickly suspected that the plaintext of message No. 173 (SIPVX) might be related to the contents of the still-unbroken MVUEH.

Crib, simulator and a Bombe

After trying many approaches, GPT-6 Astra settled on a classic cryptanalytic trick — the probable word, or crib. Here it used the place name ROSENOW ROSENOW, which appeared twice.

To carry the attack through, the system wrote the necessary software itself: in Python and C++ it built an Enigma simulator and a counterpart to the wartime Bombe, the device used at Bletchley Park to search for machine settings. Running a thorough search with the ROSENOW crib, the model ultimately arrived at the correct key and recovered plaintext of the MVUEH message.

What comes next

The researchers admit they are still combing through GPT-6 Astra's logs to work out exactly how the break was executed — and, they say, keep uncovering fresh, remarkable details. The episode matters less as a cryptographic feat (Enigma has long since stopped being considered secure) than as an example of a model assembling the entire solution — from choosing the target to writing working tools — with no step-by-step prompting.

Source: Schneier on Security